ThesisPDF Available

Defect Detection and Condition Assessment of Adhesively-Bonded Joints Using Acoustic Emission Techniques

Authors:

Abstract and Figures

The aim of this study is to investigate the application of acoustic emission (AE) techniques to the defect-detection and monitoring of adhesively-bonded joints. Pencil Lead Breaks (PLBs) have been used as a simulated AE source to experimentally investigate the characteristics of AE wave propagation in adhesively-bonded joints, and have been combined with Artificial Neural Networks (ANNs) to provide a novel method of defect detection and sizing. Modal AE analysis has been applied to destructive testing of adhesively-bonded specimens as a novel method to differentiate between fracture- modes. Dynamic Finite Element Analysis (FEA) has been utilised to simulate the AE generation and propagation to further investigate the findings of the experimental studies and to assess the applicability of the findings over a broader range of conditions than could be achieved experimentally. PLB tests have been conducted on large (500mm x 500mm x 1mm) aluminium sheet specimens to identify the effects of an adhesive layer on AE wave-propagation. Three specimens were considered; a single sheet, two sheets placed together without adhesive, and an adhesively bonded specimen. The simulated AE source is applied to the specimens at varying propagation distances and orientations. The acquired signals are processed using wavelet-transforms to explore time-frequency features, and compared with modified group-velocity curves based on the Rayleigh-Lamb equations to allow identification of wave-modes and edge-reflections. The effects of propagation-distance and source orientation are investigated while comparison is made between the three specimens. PLB tests were also used to investigate the effect of, and to detect and size void-type adhesive defects. Defect-free specimens were used for reference, and specimens with two different void sizes were tested. The PLB source was used to generate simulated AE which would propagate through the defect region and then be recorded with the AE system. Four configurations were tested to assess the effects of source-sensor propagation distance and source and sensor proximity to the defect. Typical AE parameters of peak amplitude, rise time, decay time, duration, number of counts and AE energy were investigated. Frequency analyses by Fast Fourier Transformation (FFT), partial powers and wavelet-transform (WT) were also implemented. Artificial-Neural-Networks (ANNs), using the raw Time-Domain signal as an input, were successfully trained and tested to differentiate between the specimen-types tested and to estimate the defect sizes. AE-instrumented Double Cantilever-Beam (Mode-I fracture) and Lap-Shear (Mode-II fracture) tests were conducted on similar adhesively-bonded aluminium specimens. Linear source location was used to identify the source-to-sensor propagation distance of each recorded hit, theoretical dispersion-curves were used to identify regions of the signal corresponding to the symmetric and asymmetric wave-modes, and peak wavelet-transform coefficients for the wave-modes were compared between the two fracture-modes and assessed as an indicator of fracture-mode. It was concluded that there is a relationship between the fracture-mode and the generated wavemodes, with Mode-II fracture typically generating a relatively greater symmetric wave-mode than Mode-I fracture. Dynamic FEA was used to replicate both the PLB tests and the destructive tests, and to investigate the effects of a range of parameters which could not all be practically varied in experimental work. Adhesive Young’s modulus (representative of different adhesive types), adhesive-layer thickness, and adhesive void size were varied in the simulated PLB tests. FEA was also used to investigate the effects of fracture-mode on the generated acoustic emissions in simulated mixed mode-bending tests, conducted over a range of mode-mixities. The FEA results were found to corroborate the results of the experimental work and support a relationship between fracture-mode and generated wave-modes. It was also identified that a variety of other parameters may also affect the wave-modes, and thus need to be considered to achieve effective use of modal-analysis to differentiate between fracture-modes.
Content may be subject to copyright.
Defect Detection and Condition Assessment of
Adhesively-Bonded Joints Using Acoustic Emission
Techniques
Author: Alasdair Ruairidh Crawford
Supervisors: Dr. Mohamad Ghazi Droubi & Dr. Nadimul Haque Faisal
August 2021
A thesis submitted in partial fulfilment of the requirements of the Robert Gordon
University for the degree of Doctor of Philosophy
Acknowledgements
My sincere thanks go to:
Dr. Ghazi Droubi and Dr. Nadimul Faisal for their supervision throughout this project.
Dr. Anil Prathuru for his guidance when I first started this project and for many fruitful
discussions since then.
All of the admin. staff in the School of Engineering, particularly Dr. Rosslyn Shanks and
Kirsty Stevenson.
All of the technicians who have helped with my experiments and provided great advice. In
particular thanks to Allan MacPherson for teaching me the ’correct’ pronunciation of the
term ”PhD”...
All of the many great friends I have made in my time at Robert Gordon University (RGU).
You have made my time here so much more interesting, and I may have actually learnt more
about international cuisine than I have about acoustic emission.
Dr. Lev Roberts-Haritonov and the rest of the team at MMA Offshore for their support and
flexibility throughout the final year of this project.
My family for supporting me throughout the ups and downs of this project, especially Cara,
my long suffering fiancee who has put up with my whingeing without killing me, yet...
Abstract
The aim of this study is to investigate the application of acoustic emission (AE) techniques to
the defect-detection and monitoring of adhesively-bonded joints. Pencil Lead Breaks (PLBs) have
been used as a simulated AE source to experimentally investigate the characteristics of AE wave-
propagation in adhesively-bonded joints, and have been combined with Artificial Neural Networks
(ANNs) to provide a novel method of defect detection and sizing. Modal AE analysis has been
applied to destructive testing of adhesively-bonded specimens as a novel method to differentiate
between fracture-modes. Dynamic Finite Element Analysis (FEA) has been utilised to simulate
the AE generation and propagation to further investigate the findings of the experimental studies
and to assess the applicability of the findings over a broader range of conditions than could be
achieved experimentally.
PLB tests have been conducted on large (500mm x500mm x1mm) aluminium sheet spec-
imens to identify the effects of an adhesive layer on AE wave-propagation. Three specimens
were considered; a single sheet, two sheets placed together without adhesive, and an adhesively-
bonded specimen. The simulated AE source is applied to the specimens at varying propagation-
distances and orientations. The acquired signals are processed using wavelet-transforms to ex-
plore time-frequency features, and compared with modified group-velocity curves based on the
Rayleigh-Lamb equations to allow identification of wave-modes and edge-reflections. The effects
of propagation-distance and source orientation are investigated while comparison is made between
the three specimens.
PLB tests were also used to investigate the effect of, and to detect and size void-type adhesive
defects. Defect-free specimens were used for reference, and specimens with two different void
sizes were tested. The PLB source was used to generate simulated AE which would propagate
through the defect region and then be recorded with the AE system. Four configurations were
tested to assess the effects of source-sensor propagation distance and source and sensor proximity
to the defect. Typical AE parameters of peak amplitude, rise time, decay time, duration, number
of counts and AE energy were investigated. Frequency analyses by Fast Fourier Transformation
(FFT), partial powers and wavelet-transform (WT) were also implemented. Artificial-Neural-
Networks (ANNs), using the raw Time-Domain signal as an input, were successfully trained and
tested to differentiate between the specimen-types tested and to estimate the defect sizes.
AE-instrumented Double Cantilever-Beam (Mode-I fracture) and Lap-Shear (Mode-II frac-
ture) tests were conducted on similar adhesively-bonded aluminium specimens. Linear source-
location was used to identify the source-to-sensor propagation distance of each recorded hit, the-
oretical dispersion-curves were used to identify regions of the signal corresponding to the sym-
metric and asymmetric wave-modes, and peak wavelet-transform coefficients for the wave-modes
were compared between the two fracture-modes and assessed as an indicator of fracture-mode.
It was concluded that there is a relationship between the fracture-mode and the generated wave-
modes, with Mode-II fracture typically generating a relatively greater symmetric wave-mode than
Mode-I fracture.
Dynamic FEA was used to replicate both the PLB tests and the destructive tests, and to inves-
tigate the effects of a range of parameters which could not all be practically varied in experimen-
tal work. Adhesive Young’s modulus (representative of different adhesive types), adhesive-layer
thickness, and adhesive void size were varied in the simulated PLB tests. FEA was also used to
investigate the effects of fracture-mode on the generated acoustic emissions in simulated mixed-
mode-bending tests, conducted over a range of mode-mixities. The FEA results were found to
corroborate the results of the experimental work and support a relationship between fracture-mode
and generated wave-modes. It was also identified that a variety of other parameters may also af-
fect the wave-modes, and thus need to be considered to achieve effective use of modal-analysis to
differentiate between fracture-modes.
CONTENTS
1 Research Context 1
1.1 Research Methodology and Objectives . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Contribution to Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 ThesisOutline.................................... 4
2 Literature Review 6
2.1 Adhesive-Bonding ................................. 6
2.1.1 AdhesiveTypes ............................... 7
2.1.2 Preparation, Application, and Curing of Adhesives . . . . . . . . . . . . . 7
2.1.3 Adhesive Defects and Failures . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.4 NDT of Adhesively-Bonded Joints . . . . . . . . . . . . . . . . . . . . . 10
2.2 AcousticEmission ................................. 11
2.2.1 Introduction................................. 11
2.2.2 Advantages and Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Wave-Propagation.............................. 12
2.2.4 Attenuation ................................. 18
2.2.5 Time-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.6 Frequency-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.7 Time-Frequency-Domain Analysis . . . . . . . . . . . . . . . . . . . . . 26
2.2.8 ModalAnalysis ............................... 26
2.2.9 Source-Location............................... 27
2.3 Artificial Intelligence in AE Analysis . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Application of AE to Adhesive-Bonds . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Finite Element Simulation of Acoustic Emission . . . . . . . . . . . . . . . . . . 38
2.6 Summary ...................................... 48
3 Pencil-Lead-Break-based AE Tests 49
3.1 Introduction..................................... 49
3.2 Materials ...................................... 50
3.2.1 Defect-FreeTests .............................. 50
3.2.2 Void-Type-Defect Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 AEInstrumentation................................. 51
3.4 ExperimentalProcedure............................... 52
3.4.1 Defect-FreeTests .............................. 52
3.4.2 Void-Type-Defect Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 AE Signal-Processing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.1 Wavelet-Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.2 Lamb wave Dispersion-Curves . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.3 Edge-Reections .............................. 56
3.5.4 Wavelet-Transform Example . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Results of Defect-Free Simulated-Source Tests . . . . . . . . . . . . . . . . . . 60
3.6.1 ModalAnalysis ............................... 60
3.6.2 Attenuation ................................. 66
3.6.3 Frequency-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . 67
3.6.4 Effects of Source-Orientation . . . . . . . . . . . . . . . . . . . . . . . . 72
i
3.7 Results of Simulated-Source Tests featuring Voids . . . . . . . . . . . . . . . . . 75
3.7.1 Time-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.7.2 Pattern Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.8 Discussion...................................... 88
3.8.1 ModalAnalysis ............................... 88
3.8.2 Attenuation ................................. 89
3.8.3 Frequency-Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . . 89
3.8.4 The Effects of Void-Type Defects . . . . . . . . . . . . . . . . . . . . . . 90
3.9 Summary ...................................... 92
4 AE-Instrumented Destructive Tests 93
4.1 Introduction..................................... 93
4.2 Materials ...................................... 94
4.3 ExperimentalProcedure............................... 96
4.4 AE Signal-Processing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.5 Results of Destructive Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.5.1 Load/Displacement Results . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.5.2 Failure Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.5.3 Relationship between AE and Loading . . . . . . . . . . . . . . . . . . . 103
4.5.4 AE Source-Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.5.5 Parametric Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.5.6 ModalAEResults.............................. 114
4.6 Discussion...................................... 114
4.7 Summary ...................................... 116
5 FEA of PLB tests 117
5.1 Introduction..................................... 117
5.2 Geometry ...................................... 118
5.3 Materials ...................................... 119
5.4 Boundary Conditions and Load Steps . . . . . . . . . . . . . . . . . . . . . . . 119
5.5 Source........................................ 120
5.6 Sensor........................................ 121
5.7 Mesh ........................................ 122
5.8 Solver........................................ 124
5.9 Post-processing ................................... 124
5.10 Effects of Adhesive Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.10.1 Effects of Adhesive Young’s Modulus . . . . . . . . . . . . . . . . . . . 126
5.10.2 Effects of Adhesive Thickness . . . . . . . . . . . . . . . . . . . . . . . 126
5.11 Effects of Adhesive Defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.12FEAResults..................................... 127
5.12.1ModelValidation .............................. 127
5.12.2 Effects of Adhesive Young’s Modulus . . . . . . . . . . . . . . . . . . . 135
5.12.3 Effects of Adhesive Thickness . . . . . . . . . . . . . . . . . . . . . . . 140
5.12.4 Effects of Adhesive Defects . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.13Discussion...................................... 153
5.14Summary ...................................... 154
6 FEA of Destructive Tests 156
ii
6.1 Introduction..................................... 156
6.2 Geometry ...................................... 156
6.3 Mesh ........................................ 157
6.4 Materials ...................................... 158
6.5 Boundary Conditions and Load Steps . . . . . . . . . . . . . . . . . . . . . . . 158
6.6 Source........................................ 162
6.7 Sensor........................................ 162
6.8 Post-processing ................................... 163
6.9 Results........................................ 163
6.10Discussion...................................... 174
6.11Summary ...................................... 176
7 Conclusions 177
7.1 Simulated-SourceTests............................... 178
7.2 Simulated-Source Void Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.3 DestructiveTests .................................. 180
7.4 FEAofPLBTests.................................. 180
7.5 FEAofDestructiveTests .............................. 182
7.6 FutureWork..................................... 183
Appendices 185
A Sensor Calibration Certificates 186
Annotated Bibliography 189
iii
Nomenclature
A0Fundamental anti-symmetric Lamb mode
A1,2,3...n Higher order anti-symmetric Lamb modes
AE Acoustic Emission
ASTM American Society for Testing and Materials
CWT Continuous Wavelet Transform
CFL Courant-Friedrichs-Lewy (Stability Criteria)
clLongitudinal wave velocity
cpPhase velocity
ctShear wave velocity
DAQ Data Acquisition Card
DCB Double-Cantilever-Beam
FEA Finite Element Analysis
FEM Finite Element Method
FFT Fast Fourier Transform
2D-FFT Two-Dimensional Fast Fourier Transform
hHalf the sheet thickness
kHz Kilo Hertz
LS Lap-Shear
mm Millimetres
MHz Mega Hertz
MMB Mixed-Mode-Bending
NDT Non-Destructive Testing
NI National Instruments
PLB Pencil Lead Break
PSD Power Spectral Density
PTFE Polytetrafluoroethylene
S0Fundamental symmetric Lamb mode
S1,2,3...n Higher order symmetric Lamb modes
SCU Signal Conditioning Unit
SEM Scanning Electron Microscopy
SHM Structural Health Monitoring
STFT Short-Time Fourier Transform
WT Wavelet Transform
gGroup-velocity
kWave number
ωAngular frequency
iv
Publications from this Thesis
Published:
1. Crawford, Alasdair, Mohamad Ghazi Droubi, and Nadimul Haque Faisal. ”Analysis of
acoustic emission propagation in metal-to-metal adhesively-bonded joints.” Journal of Non-
destructive Evaluation 37.2 2018: 33.
2. Crawford, Alasdair R., Mohamad Ghazi Droubi, and Nadimul Haque Faisal. ”Modal acous-
tic emission analysis of mode-I and mode-II fracture of adhesively-bonded joints.” presented
at EWGAE 33rd European Conference on Acoustic Emission Testing, Senlis, France, Sept.
12-14, 2018.
Under preparation:
1. Crawford, Alasdair R., Mohamad Ghazi Droubi, and Nadimul Haque Faisal. ”Detection and
Sizing of Adhesive Defects by use of a Hsu-Nielsen Source”. Currently under preparation
for publication.
2. Crawford, Alasdair R., Mohamad Ghazi Droubi, and Nadimul Haque Faisal. ”Acoustic
Emission Characteristics of Adhesively-Bonded Specimens: An FEM Study”. Currently
under preparation for publication.
3. Crawford, Alasdair R., Mohamad Ghazi Droubi, and Nadimul Haque Faisal. ”Finite El-
ement Simulation of Acoustic Emission from Adhesive Bond Failure under Mixed-Mode
Bending”. Currently under preparation for publication.
v
Chapter 1 Research Context
Chapter 1
Research Context
Structural adhesives have rapidly gained popularity as a joining method in a variety of sectors,
particularly the aerospace and renewable energy sectors, due to their various advantages over
more traditional joining methods. They can however be subject to the introduction of a variety of
defects, during both manufacture and service, which can ultimately lead to catastrophic failure.
The safe use of adhesives in critical applications is therefore dependent on strict quality control,
followed by regular inspection or monitoring of the adhesives’ condition to ensure any degradation
of the joint is detected early enough to prevent catastrophic failure.
Many techniques have been developed to identify the presence of defects in adhesively-bonded
specimens. These techniques have included assorted ultrasonic scanning methods [1], guided
Lamb waves [2, 3], acousto-ultrasonic methods, infrared-thermography [1] and radiography [4],
and have had varying levels of success in identifying porosity, voids, disbonds and various inter-
facial defects and degradation. While, these methods have all been proven capable of identifying
certain defects, and thus inferring the potential strength of a joint, or lack thereof, no method
currently exists which can directly determine the bond strength of a joint [4]. This factor, combined
with the potential for environmental degradation, occurring due to moisture ingress, chemical
exposure, temperature or fatigue, can lead to a need for in-service testing or monitoring to ensure
bond integrity [5, 6].
While many of the aforementioned techniques can be applied to structures in service, the need
for access to the joints for inspection can be inconvenient, or in some cases even impossible due to
the operating environment. For example, consider the blades of a wind turbine. These are typically
constructed as either a ”one-piece” design, in which a stiffening spar is adhesively bonded into a
one-piece shell, or a two-piece design, in which the two halves of the blade are adhesively bonded
together, along with internal shear webs, resulting in the use of as much as 400 kg of adhesive for
a typical 42 mblade [7]. Inspection of these blades requires stopping the turbine, thus resulting in
expensive downtime, and then typically the use of rope-access technicians descending the blade
to carry out the inspection, a potentially dangerous operation which is reliant on suitable weather
conditions. Likewise, adhesive joints used in the construction of aircraft will only be accessible
during downtime once the plane has landed, meaning that for inspection the plane must be taken
out of service. A technique that therefore lends itself to long-term monitoring of adhesive joints in
service, is acoustic emission (AE) testing. AE sensors can be retro-fitted to an existing structure
for temporary monitoring, or integrated into the structure during construction, and once fitted can
be monitored remotely, removing, or significantly reducing, any subsequent need for direct access
1 of 199
Chapter 1 Research Context
to the joints. Due to the costs and added complexity of this approach, this is not a universal
solution, but in scenarios where the costs of downtime and manual inspection will outweigh the
cost of an AE system, or situations in which the consequences and costs of failure are high, AE
can provide a suitable approach. As the technology matures, it can however be assumed that the
cost of such systems is likely to fall, meaning that this will become an increasingly more viable
approach in the future. Acoustic emission is not typically used to identify the presence of defects
within a structure, but is capable of detecting their growth by detecting the propagation of elastic
waves generated by the sudden re-distributions of stress which occur from defect growth. Using a
suitably selected sensor array, AE can monitor a large area with a very small number of sensors,
and can continuously monitor the condition of an entire structure, as opposed to scanning it region
by region as is typical with most NDT methods.
Further understanding of the AE sources however forms only part of the challenge of AE test-
ing of adhesive joints, as the process of wave-propagation from source to sensor also contributes
significantly to the signal which is recorded. The presence of an adhesive layer along this prop-
agation path will therefore contribute significantly to what is recorded, and thus needs to be well
understood for quantitative AE testing to become a possibility. Previous works investigating AE
testing of adhesives typically have not considered this aspect in great detail, but have highlighted
its importance. In a study of Mode-I and -II fracture by Droubi et al. [8] it was observed that
the frequency content of the acoustic emissions varied significantly as the crack-tip progressed to-
wards the sensor. While the source of variation could not be confirmed in the study, it is believed
to be due to the characteristics of the AE propagation due to the adhesive layer, as opposed to
being a feature of the source. Some other studies, such as that by Prathuru [9] have considered
the effects of bond quality on wave-propagation, and have even used AE-based techniques for the
detection of defects. These studies have however been restricted to small coupon-type specimens,
with little investigation of larger specimens as may exist in some industrial applications.
Significantly more work has been done regarding wave-propagation in adhesives using ul-
trasound methods rather than acoustic emission, such as the investigations by Heller et al. [2].
A lot of the findings of the ultrasonic investigations, particularly those focusing on Lamb wave
propagation, can be applied to AE applications, but the studies typically focus on a significantly
higher frequency-range than is used in AE. It is therefore felt to be important to carry out fur-
ther investigation of wave-propagation in bonded specimens using typical AE sources and sensing
equipment.
1.1 Research Methodology and Objectives
While there have been a variety of studies utilising closely-related techniques, such as ultra-
sound [2] or 3D laser-vibrometry [10], to the best of the author’s knowledge, there has been little
systematic investigation of the effects of adhesive bonding on the propagation of acoustic emission
in large scale specimens. There have also been no studies conducted investigating the relationship
between the fracture-mode of adhesive joints and the different Lamb wave modes generated by
them. A series of Pencil-Lead-Break (PLB) [11] and fracture-based experiments and a variety of
dynamic FEA simulations have therefore been conducted to develop a greater understanding of
the effects of adhesive bond status on AE propagation and to assess the possibility of using modal
AE analysis techniques to differentiate between fracture-modes in adhesively-bonded joints. An
understanding of the effects of an adhesive layer on wave-propagation is critical for correct inter-
2 of 199
Chapter 1 Research Context
pretation of any AE data resulting from failure of an adhesive, or even for other unrelated sources,
which result in propagation-path through an adhesive bond. It has been chosen to investigate
the application of modal AE analysis to the study of adhesive fracture-modes, as the study of
fracture-modes can be critical to the monitoring of adhesives due to the significant differences in
load-bearing capabilities between the different modes. Modal analysis has shown great success in
a variety of other applications, but has not previously been applied to the differentiation between
adhesive fracture-modes. Dynamic finite element simulations were conducted to investigate both
the effects of the adhesive-layer on wave-propagation and to investigate the effects of fracture-
mode. The use of simulations allows for a much better controlled environment than experiments,
allowing the effects of the parameters under investigation to be completely isolated for clearer
analysis of their effects. Once a basic model has been implemented, simulations also provide an
efficient way to investigate a range of variables which may not be possible or practical to inves-
tigate experimentally due to time or cost constraints. The main objectives of this project were
therefore:
1. To develop an understanding of the effects of adhesive bond status on the propagation char-
acteristics of acoustic emission in relatively large-scale test specimens by use of a standard
PLB source. The findings of this part of the study were subsequently used to inform the
design of experiments and analysis methods for the rest of this project.
2. To investigate the effects of bonding defects on AE propagation, and to develop AE-based
methods of defect detection using a standard PLB source.
3. To experimentally investigate the relationship between the fracture-modes of adhesively-
bonded joints and the generated acoustic emissions, specifically with respect to the Lamb
wave modes generated. The primary aim of this being to investigate the potential application
of Modal-AE analysis to differentiate between fracture-modes.
4. To develop a dynamic finite element model to investigate the effects of the parameters of
adhesive layers on AE propagation. The use of the finite element approach, in addition to
the experimental work, allows a greater degree of control of the test variables and provides
an efficient way to investigate a large range of parameters.
5. To develop a dynamic finite element model to investigate the relationships between fracture-
mode and acoustic emission through dynamic finite element analysis. The development of
a finite element model will expand on the experimental findings and will again provide an
efficient way to investigate a wide range of parameters in a much more controlled manner
than is possible experimentally.
1.2 Contribution to Knowledge
The initial contribution to knowledge from this work lies in the systematic experimental study
of AE propagation in relatively large scale adhesively-bonded specimens. The results of which
were then used to provide confidence in the techniques in the following study of adhesive frac-
ture. The study of wave-propagation in adhesively-bonded joints has demonstrated the effects of
the adhesive layer on the wave-modes generated, the frequency content, amplitude, energy and
other typical AE parameters. The results correlate well with those previously reported in studies
utilising ultrasound techniques, but are unique in terms of the use of the AE equipment and the
3 of 199
Chapter 1 Research Context
accompanying focus on a lower frequency-range than is considered using ultrasound. The use of
a PLB source to detect void-type adhesive defects is contributing to expanding on previous work
conducted at Robert Gordon University, in which this method has been used to investigate other
types of adhesive defects, as well as defects in other fields, such as weld defects. The introduction
of basic artificial neural networks (ANNs) to this approach yields a significant improvement in
terms of accuracy, and in terms of simplifying the interpretation of results.
The main novelty of this work lies in the application of modal acoustic emission analysis to
the differentiation between fracture-modes of adhesively-bonded joints. Modal AE analysis is
not believed to have previously been applied to the fracture of adhesively-bonded joints. In this
study, through both experimental work and simulation, it is shown that the fracture-mode does
significantly affect the wave-modes generated. It is however also seen that variation of other
parameters has a similar effect on the wave-modes, and thus simple modal analysis alone does not
provide a robust classifier of fracture-mode.
Further novelty lies in the FEA simulation of AE generation and propagation in adhesively-
bonded joints. This is believed to be the first instance of the use of dynamic finite element analysis
to simulate AE generation and propagation in adhesively-bonded joints. The use of this method
has allowed controlled investigation of some of the parameters affecting wave-propagation, and
has thus allowed the generalisability of previous experimental results to be investigated. The use
of simulation has also allowed analysis of the true effects of fracture-mode to be investigated, with
all other parameters remaining constant, a feat which is not typically possible with an experimental
set-up due to the variation in real-world bond quality.
1.3 Thesis Outline
This thesis is structured in 7 chapters, a summary of their content is provided below:
1. Introduction
This chapter introduces the general topic of acoustic emission testing of adhesively-bonded
joints and summarises the current state of research in this area. It also outlines the objectives
of this work and the novelty and contribution of the research.
2. Literature Review
This chapter provides an introduction to, and critical review of, the current state-of-the-art
in areas critical to this work. First, an overview of the advantages of, and problems faced
in, adhesive bonding is given. This is followed by a review of acoustic emission, in terms of
basic working theory and equipment and also current analysis techniques. A more detailed
review of the application of acoustic emission techniques to investigation of adhesively-
bonded joints is then given, highlighting the gaps in knowledge this thesis aims to fill. This
is followed by a review of the relatively young and rapidly-developing field of finite element
simulation of acoustic emission.
3. Pencil-Lead-Break based AE Tests
This chapter describes a series of experiments using a pencil-lead-break source on various
bonded, un-bonded and defective specimens to establish the effects of the bonded layer on
AE propagation. This chapter includes details of the experimental set-up, signal processing,
and analysis and discussion of the results obtained.
4 of 199
Chapter 1 Research Context
4. AE-Instrumented Destructive Tests
This chapter contains details of experimental work conducted to investigate the potential use
of modal AE analysis to differentiate between Mode-I (Crack-opening) and Mode-II (Shear)
fracture. It includes the experimental set-up, signal-processing methodology and analysis
and discussion of the results.
5. FEA of PLB Tests
This chapter describes the finite element simulations of the pencil-lead-break tests. This
includes the development of geometry, boundary conditions, meshing, simulation settings
and post-processing methods which were utilised. It also includes validation of the model
against theoretical and experimental results, and analysis and discussion of the results ob-
tained.
6. FEA of Destructive Tests
This chapter details dynamic finite element simulations of mixed-mode-bending tests devel-
oped to further investigate the differences in the wave-modes excited by different fracture-
modes. It describes the development of geometry, boundary conditions, meshing, simulation
settings and post-processing methods, and includes analysis and discussion of the results.
7. Conclusions
This chapter summarises the most significant findings of the study, as detailed in the previ-
ous chapters. Recommendations for future work and for the application of this research are
also provided.
5 of 199
Chapter 2 Literature Review
Chapter 2
Literature Review
This chapter provides an overview of pertinent literature and background theory relating to the
topics of adhesive-bonding, Acoustic Emission theory and application, and Dynamic Finite Ele-
ment Analysis of Acoustic Emission. This chapter highlights gaps in the current knowledge and
abilities in these fields, and introduces a number of concepts and techniques used throughout this
project.
2.1 Adhesive-Bonding
Since the introduction of synthetic polymers in the early 1900s adhesives have increasingly be-
come used in the place of more traditional mechanical fasteners, such as bolts and rivets. When
used correctly they are capable of providing a preferable stress distribution across the area of a
joint (not an entirely uniform stress distribution as sometimes stated), allowing greater joint stiff-
ness and higher loading bearing. They also provide a number of other advantages such as being
lighter than equivalent mechanical joints, having useful damping properties, offering good corro-
sion resistance and being suitable for joining dissimilar materials [12]. Due to these properties they
are also increasingly being used in place of welding; in this context they are also advantageous
as they can avoid heat-induced sensitisation, deformation or burn-through of the materials being
joined. Due to their improved stress distribution, adhesives are the fastener of choice for com-
posite materials which can be unsuitable for the high bearing stresses introduced by mechanical
fasteners [12].
These advantages have led to adhesive bonding being used in many industries. In the aerospace
industry it is used extensively for the bonding of skins to the underlying framework, but in some
cases also for the joining of inner and outer frame members [13]. In space exploration, adhesives
have been extensively used for the construction of composite panels, created by bonding metallic
skins to honeycomb cores [14]. The rail industry uses adhesives extensively for internal finishing
parts of train carriages, but as movement is being made towards lighter, more efficient trains, their
use in more significant structural components, such as composite roof panels, is becoming more
prominent [15]. The automotive industry uses adhesives for a variety of internal trim, but also for
hem-flange bonding of panels such as doors and bonnets, a process in which one sheet of metal
is folded over the edge of another, to join the two panels together while creating a strong and
stiff edge. Another use in the automotive industry is anti-flutter bonding, in which adhesive is
6 of 199
Chapter 2 Literature Review
applied between the inner and outer metal panels to prevent rattling [16]. The marine industry
also uses adhesives extensively, particularly in the construction of small pleasure-craft and racing-
yachts. Common applications being bonding of decks to hulls, bonding bulkheads to hulls, and
the construction of composite honeycomb panels used in high-performance yachts [17]. The wind
energy industry is also heavily reliant on adhesive bonding, with composite turbine blades often
being made in two halves, which are then bonded together, or as a single outer shell, which then
has internal stiffening beams bonded into it [7].
2.1.1 Adhesive Types
An ever-growing array of adhesives exists, with new formulations being constantly developed and
tailored to suit specific applications. According to Papon [18], these can be broken down into
three main families of adhesives, though with many more sub-categories within these. These three
main families are; adhesives implemented via a physical process, pressure-sensitive adhesives, and
adhesives implemented via chemical bonding.
Adhesives implemented by physical processes, refers to adhesives which change from a liquid-
state to solid-state to form a bond by evaporation of the solvent component of the adhesive, water
diffusion, or cooling. This includes water-based, solvent-based, dispersion-based and hot-melt
adhesives. These types of adhesive are cheap and readily available, but of relatively low strength.
They are used extensively for arts and crafts, furniture production, stationary, and medicine, but
are generally not well suited for industrial use.
Pressure sensitive adhesives are visco-elastic solids which which can instantly adhere to a sur-
face with the application of pressure [19]. These are typically silicone-, polyacrylate- or polydiene-
based adhesives, which are generally supplied as self-adhesives for tapes, films or paper. These
adhesives are typically very low strength and are non-permanent or semi-permanent, but offer the
advantages of being fast and also of being repositionable [18]. Typical examples of their use are
masking-tape, sticky-notes and sticking plasters.
Adhesives implemented by chemical reaction make up the majority of structural adhesives
used in industrial applications, as well as high performance household adhesives. These adhe-
sives work by the process of polymerisation, in which individual molecules join together to form
a chain. This reaction can be initiated by the mixing of the adhesive with a catalyst (2-part adhe-
sive), the reaction of the adhesive with elements in the environment such as oxygen or moisture, or
the reaction of the adhesive with an external energy source such as heat, UV light or electromag-
netic radiation [18]. Typical examples of these are 2-part Epoxies, Aminoplasts, Phenoplasts and
Cyanoacrylates amongst others. These adhesives are available as liquids, pastes, aerosol sprays
and also pre-impregnated tapes and sheets. When used correctly, these adhesives can offer very
high-strength, permanent joints between a wide range of different materials.
2.1.2 Preparation, Application, and Curing of Adhesives
One-part adhesives, which react with the environment or with an energy source can be the sim-
plest to work with, as a single layer of the adhesive can be applied straight from the container,
without need for mixing. Two-part adhesives do however require mixing of the two parts prior
to application. This step is critical, as the correct quantities of each part must be mixed to create
a stoichiometric mixture, in which the reactants are correctly balanced for polymerisation [20].
7 of 199
Chapter 2 Literature Review
Improper measurement of each part, or inadequate mixing of the parts can result in sections of
unpolymerised adhesive, which has little to no strength. Two-part adhesives can be supplied with
each part in a separate container, which for manual measurement and mixing means the compo-
nents must be weighed to ensure the correct ratio before mixing manually with a spatula [21]. A
better solution for manual mixing is the supply of both parts in a double-barrelled syringe, ensur-
ing that the correct ratio of the parts can be easily acquired without having to weigh them. The
two parts can then be mixed manually, or in some cases a mixing nozzle can be fitted to the sy-
ringe, so the adhesive can be applied directly to the adherends from the syringe, having already
been mixed [22]. For more industrial applications in which large quantities of adhesives are being
mixed, specialised mixing machines can be used. These vary in design based on the properties and
quantity of the adhesive under preparation, but typically consist of a fixed mixing drum with some
form of rotating mixing blade, agitator, or kneader. Centrifugal mixers in which the whole drum
rotates in the opposite direction to the agitators also exist and are advantageous due to their fast
mixing time. Vacuum mixing, the mixing of adhesives at below atmospheric pressure, is also used
for certain applications. This has the advantage of de-aeration, which reduces the chance of voids
or porosity within the adhesive, as well as being advantageous for certain adhesive formulations
which may react in an undesirable manner with oxygen or with water present in the atmosphere. In
some cases the measurement and mixing of the adhesives is done continuously using an integrated
metering and mixing system, which significantly improves workflow when compared to the batch
preparation of adhesives [22].
The most basic form of adhesive application is manual application, in which the adhesive is
applied to the adherends using some form of spatula or spreading stick, or is extruded onto the
adherends from the nozzle of a manual syringe or handgun-type applicator. This is cheap and
simple, but the quality of the joint can be greatly affected by the skill of the personnel applying the
adhesive. In larger-scale industrial applications, robotic applicators can be used to apply a bead,
or spray a jet, of adhesive onto the adherends. Depending on the application, this can be done with
either a fixed nozzle, which parts pass underneath on a conveyor, or with a nozzle mounted on a
robot arm, allowing it to move in three dimensions. While the use of automated systems does not
necessarily lead directly to a stronger joint, removing the element of human error allows joints to
be made much more consistently, and therefore makes their behaviour much more predictable [22].
2.1.3 Adhesive Defects and Failures
While adhesive-bonding does offer many advantages and has been widely adopted across these
industries, its use in safety-critical applications has been restricted by inadequate methods of
non-destructive testing [4]. As discussed in the following sections there are a variety of poten-
tially strength-reducing defects which may occur in bonds, only some of which can be readily
detected with conventional NDT methods. The consequences of failure of an adhesive joint can
vary greatly, depending on the application. In many cases it may just result in the minor inconve-
nience and cost of having to carry out repair work. At the other end of the spectrum is the Aloha
Airlines Flight 243 incident, in which the debonding of a lap-joint in the crown skin of a Boeing
737-200 ultimately led to the explosive decompression of the planes cabin, resulting in the loss of
one life, the severe injury of eight others, and the loss of the plane [13,23].
As illustrated in Figure 2.1, there are a variety of defects which may occur within an adhesively-
bonded joint which may result in reduced overall joint strength and ultimately lead to failure. The
majority of these defects are introduced during the manufacture of the joint. Porosity can be in-
8 of 199
Chapter 2 Literature Review
troduced by the entrapment of air or by the chemical reactions involved in the curing process of
the adhesive. Some extent of porosity is present in most bond-lines and only becomes problematic
when its presence is excessive or unexpected. Cracking of the adhesive may occur due to issues
in the curing process, such as thermal shrinkage, or may alternatively occur due to overloading
or fatigue during service. Poor cure of the adhesive can occur due to incorrect proportions of the
adhesive components being mixed, improper mixing of these components or insufficient thermal
exposure in thermally-activated adhesives. Voids may be introduced to a joint by air becoming
trapped during the lay-up of the joint. This can be caused by insufficient or uneven spreading of
the adhesive and also by relative movement between the adherends during cure. Surface unbonds
are a type of void located between the adherend and adhesive which generally are the result of
the adhesive being unevenly applied to only one adherend before the adherends are joined. A
zero-volume unbond, or kissing bond, is an unbond in which the adhesive makes contact with the
adherend, but does not fully adhere to it. This often results from inappropriate surface prepara-
tion or contamination. The result can vary from merely exhibiting reduced bond strength, to no
bond strength at all. This is one of the most dangerous types of defects as it is hard to detect with
standard non-destructive testing methods as there is no volume of void [4].
Figure 2.1: Potential bond-line defect types
The criticality of these defects is not only dependent on their severity, but also on their location
and environmental conditions. Adhesive-bonds do not exhibit a uniform stress distribution, as is
commonly assumed, but rather feature higher stresses around the edges than in the centre. This
makes them highly tolerant of defects existing within the low stress central region of the joint, as
has been demonstrated both theoretically and experimentally by a number of authors [4, 24,25].
There are three main fracture-modes by which an adhesive joint can fail, which may occur
individually or in combination, producing a mixed-mode failure. The main modes are illustrated
in Figure 2.2. Mode-I is characterised by crack-opening, while modes -II and -III are both shearing
modes, more specifically referred to as sliding, or in-plane shear, and tearing, or out-of-plane shear.
Joints are most commonly designed to be predominantly loaded in tensile shear, as this is
how they are strongest. Peel and cleavage loads should be avoided wherever possible, as adhesive
bonds are much more susceptible to failure under these loadings [26, 27]. While it should be min-
imised by suitable design, mixed-mode loading is however still a common occurrence for a variety
of reasons. These can include other design constraints which prevent the joint from being oriented
9 of 199
Chapter 2 Literature Review
in pure shear, varying loading orientations, deformation of the structure under load leading to vari-
ation in load orientation, manufacturing eccentricities, elastic mismatch of the adherends, thermal
mismatch of the adherends and the adhesive, or the introduction of additional unplanned loadings,
such as those from impact or collision [27]. Even a simple single-lap joint loaded in pure tension
can be subject to some bending and the introduction of peel loads at the ends of the overlap due to
the adherends being offset from each other.
The failure resulting from these loadings and the previously described defects can be described
as adhesive, cohesive or adherend failure, or any combination of these. Adhesive failure is the
separation of the adhesive from one or more of the adherends due to failure of the bonding between
them. Cohesive failure is failure occurring within the adhesive itself, generally leaving a layer of
adhesive stuck to both adherends. Adherend failure occurs when the adherend yields before the
adhesive fails.
Figure 2.2: Fracture-modes
2.1.4 NDT of Adhesively-Bonded Joints
Quality assessment and condition monitoring of adhesive joints is faced with a number of chal-
lenges. First of all there is no non-destructive method which can directly assess strength of ad-
hesion, this can only be done by destructive testing. There are a number of measurable variables
which can be interpreted as an indication of strength, or lack thereof, but are not direct measure-
ments of bond strength [4]. Methods such as ultrasound and X-ray have been well proven for
identification of voids and cracks and can be used to assess the contact area of the bond, but they
do not however assess the strength of the bond, so while they are valuable tools for identifying
certain defects they are unable to give an overall picture of bond strength. Another limiting issue
is the time taken by scanning methods such as ultrasound. In a large bonded structure, the scan-
ning of the entire bonded area can be an extremely time consuming process. This becomes further
complicated by access to the joint. In many practical situations access is only available to one
10 of 199
Chapter 2 Literature Review
side of the joint being assessed, which limits the type of testing that can be carried out [4]. The
inaccessibility of certain structures also provides a challenge to conventional inspection meth-
ods [28], although advances in robotics technology are improving this situation, with a variety
of remote controlled and autonomous inspection robots having been developed, such as the six-
legged, suction-cup footed, walker-robot developed by Herraiz et al. [29], or the suction-supported
or magnet-supported tracked crawler robots built by International Climbing Machines [30]. These
robots have been developed to climb structures such as wind turbine blades, aircraft fuselages or
ships hulls, and can be equipped with cameras, ultrasound sensors and/or AE sensors, depending
on the application.
Many techniques to assess adhesive-bond quality have been developed, with varying levels of
success and differing merits in terms of accuracy and practicality. Various ultrasonic techniques,
such as through-transmission, pulse-echo and pitch-and-catch systems are widely used throughout
industry. While they are extremely effective in certain situations, they can be limited by aspects
such as; requiring access to both sides of the bond (for through-transmission), the limited depth
that can be inspected by single-sided approaches,the inability to detect certain defects such as zero-
volume disbonds, and the necessity for sensor coupling by water jet or immersion bath, though this
can be avoided by the use of air-coupled systems [31]. These techniques are also generally reliant
on scanning of the entire area being inspected, an extremely time-consuming process for large
areas, with areas of several square metres potentially taking over an hour to scan, depending on
the desired resolution [4], though this can be improved by the use of Lamb-, or plate-waves, which
can be used to inspect a path rather than just a point. Techniques such as radiography and infrared-
thermography can inspect larger areas much faster, but radiography is largely insensitive to the
presence of adhesive unless it is combined with a metallic filler, as the density of the adherends
is generally much higher than that of the adhesive [32]. While infrared thermography provides a
similar sensitivity to near-surface defects as ultrasonic pulse-echo techniques, it is less sensitive
to deeper defects and is generally unsuitable for inspection of both thin layers and specimens
made of highly conductive materials, such as metals [4]. A variety of other techniques including
impedance, and sonic- and ultrasonic vibration based methods are also available and have their
own advantages; the majority, however, are still restricted by the time-consuming requirement of
scanning of the bond area. One technique which avoids this issue is acoustic emission.
2.2 Acoustic Emission
2.2.1 Introduction
Acoustic emission is the phenomena of transient elastic waves being generated by the sudden re-
distribution of stress within a material. The elastic waves will propagate through the material to
the object’s surface, where they can then be detected by sensors. Acoustic emission can be gen-
erated by a number of sources including; mechanical deformation, fracture, phase transformation,
corrosion, friction and magnetic processes [33]. AE differs significantly from the majority of other
NDT techniques in two respects. First of all, the signals detected by the AE system are generated
by the object which is under examination, rather than being generated by the test equipment. Sec-
ondly, AE is the investigation of dynamic processes such as the development of defects within
a specimen. AE testing is therefore not concerned with detecting the presence of defects in the
manner which other NDT techniques tend to be, but is appropriate for monitoring the initiation
and progression of defects. AE testing therefore requires some external stimulus such as a load
11 of 199
Chapter 2 Literature Review
being applied to the object under investigation. This makes AE well suited to the monitoring of
structures and systems which are in operation, as the working conditions can be enough to trigger
the initiation or development of defects and therefore cause the generation of acoustic emissions.
In other scenarios, a proof load can be applied to specimens specifically with the intent of gen-
erating AE; the loading at which the onset of AE occurs can then be used as an indicator of the
specimens condition [34].
2.2.2 Advantages and Limitations
AE testing is generally not an alternative to other NDT methods, such as ultrasound, but is a com-
plimentary method, as all of the available techniques have various strengths and weaknesses [34].
AE has the advantage of being able to cover large areas of a structure at once, and offers the op-
portunity to continuously, or semi-continuously, monitor the structure, with either permanently-
or temporarily-installed sensors. AE is also ideal to use while equipment is in operation, partic-
ularly as, once installed, there is no requirement for an operator to be present [34]. This poses a
significant advantage for situations in which an operator could not be present during operation, for
example on the wings of a plane or the blades of a wind turbine. The ability to perform source-
location is also a great advantage of the technique, as defects and damage can be located quickly
without inspecting the entire structure. Additionally, as AE detects defect propagation, as opposed
to detecting defect presence, there is no minimum physical defect size required for a defect to be
detected [35]. The fact that AE only detects propagating defects can also be advantageous in the
case of defect-tolerant structures, as it will allow defects which pose a threat to the structure to be
identified, while other more harmless defects will be ignored. It is emphasised by Hart-Smith [13]
that while adhesive bonding may offer an improved stress-distribution compared to other joining
methods, that the stress distribution is not actually uniform, as is often wrongly stated, and thus
not all defects or damage will actually cause a reduction in joint strength.
While AE has many advantages, it is not free of limitations. The fact that AE is currently a
more qualitative than quantitative technique, and cannot directly detect the size of defects, means
that its usefulness is limited when it comes to predicting the remaining lifespan of a structure, or
the necessity for repair. It is therefore best used in conjunction with other techniques which are
better suited to this aspect of NDT. The requirement for defects to be propagating for detection also
poses a disadvantage for the technique, as it means that for AE to be used, damage must be caused
to the structure, making it more of a semi-destructive testing technique than non-destructive.
2.2.3 Wave-Propagation
Waves propagating in an elastic medium can be represented by the general wave Equation 2.1,
with the use of appropriate boundary conditions [36].
2
∂t2=c22(2.1)
Where:
t: time
c: wave-velocity
12 of 199
Chapter 2 Literature Review
2: Laplacian operator in cartesian coordinates = 2
∂x2+2
∂y2+2
∂z2
: Potential function, representing two plane waves propagating in positive and negative
directions
For deformation occurring in a single direction, for example the x-direction, the general solu-
tion of the potential function is:
=f(xct) + F(x+ct)(2.2)
Where:
f: Function corresponding to wave travelling in +x direction
F: Function corresponding to wave travelling in -x direction
As previously stated, the boundary conditions determine how the elastic waves will actually
propagate, and will determine the characteristics of any recorded AE signal. Within an infinite
medium, or something that can be approximated as such (such as within the bulk of a large block
specimen), waves will propagate as a combination of longitudinal and shear waves. Within an a
medium bound by only one surface, such as close to the surface of a large block, the elastic waves
will propagate as surface waves of either the Rayleigh or Love type. In a medium bound by two
surfaces, such as a thin plate or sheet, waves will propagate in the symmetrical or asymmetrical
Lamb-modes, also referred to as the extensional or flexural Lamb-modes.
General Wave Equation
Infinite
Medium
(Bulk)
Longitudinal
Wave
Shear
Wave
Semi-infinite
Medium
(Surfaces)
Rayleigh
Wave
Love
Wave
Infinite Medium Bounded
by two Surfaces
(Sheets & Plates)
Lamb
Wave
Extensional/
Symmetric Wave
Flexural/
Asymmetric Wave
Figure 2.3: Wave-modes by media type
Within the bulk of a large specimen, AE waves consist of longitudinal and shear waves. In
a Longitudinal wave, also know as a compression, dilation, pressure, or P-wave, localised com-
pression and dilation of the material take place with the particle motion in line with the direction
of wave-propagation, as shown in figure 2.4 [37–39]. Shear waves, also referred to as transverse,
or S-waves, consist of particle motion perpendicular to the direction of wave-propagation, as per
figure 2.5.
13 of 199
Chapter 2 Literature Review
Figure 2.4: Longitudinal Wave
Figure 2.5: Shear Wave
The propagation velocities of both longitudinal and shear waves are independent of frequency,
and are governed by the following expressions [37, 39]:
For Longitudinal waves:
c1=sλ0+ 2µ
ρ(2.3)
For shear waves:
c2=rµ
ρ(2.4)
Where:
c1: Longitudinal wave-velocity
λ0: Lame’s constant = Ey
(1+v)(12v)
µ: Rigidity Modulus = Ey
2(1+v)
Ey: Young’s Modulus
14 of 199
Chapter 2 Literature Review
v: Poisson’s ratio
ρ: Density
The wavelength at a particular frequency can be calculated using the following equation [39]:
λ=c
f(2.5)
Where:
λ: Wavelength
c: wave-velocity
f: Wave frequency
Surface waves are dominant in semi-infinite mediums, in which there is a single free surface
located suitably far from other surfaces as to avoid interaction. These waves are commonly of
interest in AE due to the sensors generally being surface mounted, rather than embedded within
the bulk of a specimen. Two types of surface wave exist, the Rayleigh wave, and the Love wave. In
a Rayleigh wave particles oscillate in an orbital manner, much like typical water waves, moving in
the directions in and out of the surface, and along the surface in the direction of wave-propagation.
In contrast to this, a Love wave consists of particle oscillation in the direction perpendicular to the
direction of wave-propagation, similar to a shear wave [37].
Figure 2.6: Rayleigh Wave
Figure 2.7: Love Wave
The propagation of surface waves is slower than that of bulk waves, and the velocity of a
Rayleigh wave can be estimated as follows [39]:
15 of 199
Chapter 2 Literature Review
cr= 0.92c2(2.6)
Within a thin sheet or plate, wave-propagation becomes more complex, due to the interaction
of the waves on each of the surfaces. For relatively thick specimens, where the material is sig-
nificantly thicker than the wavelength, separate Rayleigh waves can propagate on each surface.
However, when the plate thickness is smaller than the wave-length, the interaction between the
surfaces results in the generation of Lamb waves, also known as plate waves [38].
Two fundamental modes of Lamb wave exist, the Symmetric and the Asymmetric modes.
These modes can be visualised as the Rayleigh waves on the opposing surfaces being in-phase with
each other (Symmetric) resulting in an extensional wave, or out-of-phase (Asymmetric) resulting
in a flexural wave.
Figure 2.8: Symmetric Lamb Wave
Figure 2.9: Asymmetric Lamb Wave
The Rayleigh-Lamb equations are given as Equations 2.7 and 2.8 below [40]:
Symmetric :tan(qh)
tan(ph)=4k2pq
(k2q2)2(2.7)
Asymmetric :tan(qh)
tan(ph)=(k2q2)2
4k2pq (2.8)
Where:
p2=ω2
c2
l
k2
q2=ω2
c2
t
k2
h: Half the sheet thickness
ω: Angular frequency
16 of 199
Chapter 2 Literature Review
k: Wave number
cl: Longitudinal wave-velocity
ct: Shear wave-velocity
Phase velocity: cp=ω
k
Group-velocity: g=ω
dk
Based on these equations, group-velocity curves, representing the variation in wave-propagation
velocities with frequency, can be plotted, as shown in Figure 2.10.
Figure 2.10: Group Dispersion-Curves for Symmetric and Asymmetric Type Lamb Waves (Ex-
ample for 1 mm Aluminium Plate)
The velocity of a Lamb wave is frequency, mode and thickness dependent, thus resulting in
rather complex behaviour for broadband signals, such as are typical in AE. A signal can contain
multiple modes and frequencies, thus giving the signal a variety of arrival times at a sensor, and
so requiring a more in-depth approach to signal processing. Which wave-modes are generated by
a source is dependent on the source orientation and position, this will be discussed further in a
subsequent section.
Propagating waves will, at some point, reach an interface between one material and another.
At this point, the wave may be transmitted into the adjoining medium, reflected, or may propagate
along the interface. Generally a combination of these mechanisms will occur. At interfaces be-
tween materials, or at changes in geometry, mode-conversion may also occur, in which waves will
change from one type to another. For example, in a specimen consisting of a large block with a
sheet attached to it, it is likely that Lamb-waves will form in the sheet section, but in the block sec-
tion, only bulk and surface waves will propagate, so mode conversion will occur at the change in
geometry. In the case of this project, mode conversion can occur at the interface between bonded
and un-bonded sections of a specimen.
The reflection and refraction of the elastic waves is determined by the angle of incidence and
the acoustic impedance (Z) of the two materials. Acoustic impedance is a material property that
17 of 199
Chapter 2 Literature Review
can be calculated as follows [38]:
Z=ρ×c(2.9)
Where:
ρ: Material density
c: wave-velocity
The difference in acoustic impedance between the two materials can be used to approximate
the proportion of energy which will be transmitted, and the proportion which will be reflected. For
materials with impedances Z1and Z2, the proportion of energy reflected Ercan be approximated
by:
Er=(Z1Z2)2
(Z1+Z2)2(2.10)
For materials of similar impedance, energy will be largely transmitted, while materials of
greatly differing impedance will result in a high level of reflection.
2.2.4 Attenuation
While elastic waves can travel great distances under the correct conditions, their amplitude will
reduce as the distance from the source increases. This phenomenon is called attenuation, and
can be attributed to four main mechanisms; geometric attenuation, scattering/diffraction, material
damping and dispersion.
Geometric attenuation is due to the spreading of a wave in space. In an infinite 3D space, the
wave from a localised source will propagate outwards in a spherical pattern. Due to conservation
of energy, the amount of energy contained within the wave remains constant. Thus as the radius of
the wave-front increases, and the area over which the energy is spread increases, the concentration
of energy at any given point on the wave-front must decrease. The wave amplitude (A)is therefore
proportional to the inverse of the radius (r)of the wave-front (A1/r). In a thin, sheet-
type specimen, the geometric spreading can be approximated as being 2D, with the wave-front
expanding as a cylinder of increasing radius, rather than a sphere. In this case the amplitude (A)
will reduce proportionally to the square root of the inverse of the radius (r)of the wave-front
(Ap1/r)[41, 42]. Attenuation in real-world structures is however more complex, due to
the impact of reflections which can act to reduce the effective attenuation, in all but the largest
of structures. In thin sheets or narrow rods, waves can cover significant distances with minimal
attenuation, a feature which can be extremely useful when correctly exploited for techniques such
as AE or certain types of ultrasound.
Scattering and diffraction contribute to attenuation in inhomogeneous media, where voids,
cracks, inclusions and complex internal boundaries can re-direct the propagating wave on a lo-
calised scale. Scattering occurs in a similar manner to reflection at the edge of a specimen. When
the wave reaches an internal interface with a material of a different acoustic impedance, such as
a void, or an inclusion, a proportion of the wave will be reflected back, while the remainder of
the wave will pass through the boundary into the secondary material. Diffraction of the elastic
wave can also be caused by sharp-edged internal features, such as cracks. Both scattering and
18 of 199
Chapter 2 Literature Review
diffraction can lead to increased attenuation within a specimen. Scattering and diffraction, along
with edge-reflections, result in waves propagating in different directions, which can in turn result
in interference, which can cause a further change in attenuation. Interference occurs when mul-
tiple waves interact. If the waves are in phase with each other, then the result is that of positive
interference, and the wave amplitude will be greatly increased as the waves are essentially added
together. While if the waves are out of phase, they can cause a significant reduction in amplitude,
potentially cancelling each other out.
Material damping refers to the conversion of mechanical energy related to the wave motion
to thermal energy. Hooke’s law assumes stress and strain are proportional until yield, and in-
phase, but this is only valid when the loading-rate is so slow that the deformation process may be
considered static. With higher loading-rates however, the resultant deformation will lag behind the
applied load. In the case of a cyclic loading the lag results in a hysteresis loop, as shown in Figure
2.11. The area between the loading and unloading curves represents the energy lost as heat.
Figure 2.11: Hysteresis Loop
For materials with low damping, generally those typically considered to be linear-elastic, it is
typical to quantify material damping using the dimensionless quantities of either Specific Damping
Capacity (ψ)or Logarithmic Decrement (δ), measured by either cyclic loading or free-vibration
respectively.
Specific damping capacity is defined as the ratio of energy lost per loading-cycle to maximum
total strain energy [43]:
ψ=δW
W(2.11)
Where:
δW : Energy dissipated in a cycle
W: Elastic energy stored when strain is at its maximum
When considering materials with significant visco-elastic properties, such as polymers, a more
detailed approach is necessary, particularly if attempting to simulate the material behaviour.
Multiple different models of viscoelasticity exist, all of which can be represented in terms of
springs and dashpot dampers, connected in series, parallel, or a combination of both series and
parallel.
Behaviour of a simple linear elastic material with a stiffness Ecan be represented by a single
linear elastic spring, as shown in Figure 2.12 and is governed by the constitutive Equation 2.12
19 of 199
Chapter 2 Literature Review
[43]:
ε=1
Eσ(2.12)
Figure 2.12: Linear Elastic Spring
In this case strain is instantaneous and directly proportional to the applied load. This model
makes up the elastic part of a viscoelastic model. The viscous part of the model can be represented
by a dashpot-style damper, as shown in Figure 2.13. The dashpot responds with a strain-rate
proportional to the applied stress [43]:
˙ε=1
ησ(2.13)
Assuming zero initial strain, for an applied stress σ0the strain is given by [43]:
ε=σ0
ηt(2.14)
Figure 2.13: Linear Dashpot
For as long as the stress is applied, the strain will increase linearly. Upon unloading, the strain
will remain constant, as illustrated in Figure 2.14.
Figure 2.14: Linear Dashpot Response [43]
The Maxwell model for viscoelasticity places the viscous damper and linear elastic spring in
series, as illustrated in Figure 2.15. For equilibrium, there must be uniform stress (σ2) throughout
both elements of the model, and the total strain will be comprised of strain in the spring element
(ε1) and strain in the dashpot element (ε2). This yields the following three equations [43]:
20 of 199
Chapter 2 Literature Review
ε1=1
Eσ, ˙ε1=1
ησ, ε =ε1+ε2(2.15)
which can be rearranged and combined to give the Maxwell Model:
σ+η
E˙σ=η˙ε(2.16)
The response of the Maxwell model to instantaneous loading and unloading is illustrated in
Figure 2.16. Upon application of load the spring element will be instantaneously strained, while
the dashpot will take time, with strain increasing linearly until unloading occurs. Upon unloading,
the spring will instantaneously recover, whereas there is no driving force for the dashpot to recover,
and it will thus remain strained [43].
Figure 2.15: Maxwell Model
Figure 2.16: Maxwell Model Response
In contrast to the Maxwell model, the Kelvin-Voigt model places the spring and dashpot in
parallel, ensuring uniform strain across the system. This yields the following three equations [43]:
ε=1
Eσ1,˙ε=1
ησ2, σ =σ1+σ2(2.17)
Rearranging and combining then gives the Kelvin-Voigt Model:
σ=+η˙ε(2.18)
21 of 199
Chapter 2 Literature Review
Figure 2.17: Kelvin-Voigt Model
The response of the Kelvin-Voigt model differs to the Maxwell, as on application of an instan-
taneous load, the motion of the spring is constrained by the movement of the dashpot, resulting
in no instantaneous strain. The stress is initially taken by the dashpot, which therefore controls
the initial strain-rate. As the dashpot starts to move and the spring extends, the contribution of
the spring element increases until the entire load is taken by the spring and the strain reaches a
maximum. This gives a curve with reducing strain-rate throughout the stroke, as opposed to the
linear strain-rate of the Maxwell model. Upon unloading, the spring forces motion of the dashpot,
returning the system to zero strain. As the initial return is governed by the spring, with an increas-
ing contribution from the damper throughout the stroke, the relaxation of the Kelvin-Voigt model
also follows a curve of decreasing strain-rate [43]. This response is illustrated in Figure 2.18
Figure 2.18: Kelvin-Voigt Response
To more accurately model the behaviour of real viscoelastic materials, these models can be
combined into generalised models, which contain multiple Maxwell or Kelvin-Voigt units com-
bined in either series or parallel. The Generalised Maxwell Model (Figure 2.19) consists of multi-
ple Maxwell units connected in parallel with a spring and damper, while the Generalised Kelvin-
Voigt model (Figure 2.20) consists of multiple Kelvin-Voigt units connected in series with a spring
and damper [43].
22 of 199
Chapter 2 Literature Review
Figure 2.19: Generalised Maxwell Model
Figure 2.20: Generalised Kelvin Voigt Model
The Generalised Maxwell model is well suited to representing the behaviour of solids as the
inclusion of the spring and damper in parallel result in no instantaneous strain, and does not result
in residual stress upon unloading, whereas the generalised Kelvin-Voigt model is better suited
to modelling a fluid-type response as the isolated spring and damper connected in series allows
for instantaneous displacement and residual strain. Increasing the number of elements within a
generalised model can greatly increase the accuracy with which it represents a material behaviour,
but determining suitable parameters for such a model becomes incredibly challenging, and in many
cases impractical [43].
Attenuation by dispersion occurs by the temporal and spacial separation of different wave-
modes, and different frequency-components within a single wave-mode, due to their varying ve-
locities [42]. Close to the source, the wave-modes will all ”overlap” resulting in a wave of high
amplitude and energy. Further from the source, the faster waves will have moved ahead of the
slower modes, eventually completely separating from them. The result being a signal of lower
amplitude, but longer duration.
23 of 199
Chapter 2 Literature Review
2.2.5 Time-Domain Analysis
Time-domain analysis forms a large part of traditional AE analysis. It is the simplest type of
analysis in that parameters can be easily read or calculated from the raw AE signal as recorded.
While more complex parameters and analysis methods now exist, time-domain parameters remain
very useful and are widely used. The typical AE time-domain parameters are summarised below
and are illustrated in Figure 2.21. An AE Hit, or Event, is a signal which exceeds a pre-set
threshold value. Methodologies of setting the threshold level vary, though it is recommended to
be 3dB to 4dB above the noise floor [44]. When a hit occurs, the signal, or certain parameters of
it, will be saved and used for analysis. The accumulated number of hits or rate of hits can also
be utilised in the testing of a structure. The signal duration is the time interval between the first
and last threshold-crossings of a hit. The number of counts is the number of times within the
duration that the signal exceeds the threshold. The Peak Amplitude is the peak voltage recorded
within the hit. Amplitudes are typically expressed in decibel scale where 1V at the sensor is 0
dB AE, though they can also be expressed in terms of voltage. Rise-time is the time interval
between the hit first exceeding the threshold-crossing and attaining its peak amplitude. Decay
Time is the time interval between the hit attaining its peak amplitude and the last point at which
the signal exceeds the threshold. Slight variations of the definition of AE energy exist depending
on the equipment supplier but definitions are generally based on energy being the area under the
rectified signal envelope, or in some cases the area under the rectified signal envelope but above
the threshold [34]. Throughout this work the energy has been calculated independent of the signal
threshold.
Figure 2.21: Time-Domain Parameters
24 of 199
Chapter 2 Literature Review
2.2.6 Frequency-Domain Analysis
Frequency-domain analysis is widely used for differentiation between, and identification of, AE
sources. A variety of frequency-based parameters exist, which each have various merits. Care
should however be taken in the application of frequency-based techniques as the recorded fre-
quency spectra will be affected greatly by the frequency response of the sensors and other acqui-
sition equipment used, not just by the AE source being recorded [44].
A Fourier transform of the AE signal generates a frequency spectrum and is a good way to
visualise the frequency content of a signal. Conducting a Fourier transform also allows features
such peak-frequencies to be identified. Obtaining the the frequency spectrum can give a good
overall image of the signal, but analysing and manipulating a frequency spectrum consisting of
many data points can be cumbersome and computationally expensive. Other methods are therefore
often used to summarise the frequency spectrum.
Partial powers act as a condensed frequency spectrum, by summarising the energy in a few
key frequency bands. The typical process used to obtain partial powers is to filter the signal into a
few different frequency bands, and then to calculate the energy in each of the filtered signals [44].
The frequency bands can be chosen either as regular intervals in the frequency-domain (often 100
kHz bands), or chosen to correspond to peaks identified from the frequency spectrum.
The peak-frequency is the frequency at which the frequency spectrum calculated by Fourier
transform reaches its highest value [34, 44]. The simplicity of a single value to summarise the
frequency content can be advantageous, however, for a lot of sensors the resonant-frequency of
the sensor will dominate the signal, resulting in the same peak-frequency occurring for multiple
different AE sources with different frequencies.
The frequency-centroid is the weighted mean of the frequency spectra, calculated from the
Fourier-transform using the magnitudes of the frequencies as their weightings [34, 44]. This pa-
rameter summarises the frequency spectra into a single value without being dominated by the
effects of a single peak and is thus capable of capturing subtle changes in the frequency spectra
which may not be apparent in peak-frequency analysis.
The weighted peak-frequency is the root of the product of peak-frequency and frequency-
centroid, thus combining the advantages of peak-frequency and frequency-centroid into a single
parameter [44].
Frequency domain analysis has a variety of potential applications, but has been predominantly
used to differentiate between different source-mechanisms, particularly in the failure of compos-
ites. Fourier transforms, partial powers, peak-, centroid-, and weighted-peak-frequencies were
all successfully utilised by Njuhovic et al. to differentiate between the failure-modes of matrix-
cracking, interface-failure, fibre-breakage and interphase-failure in metallised GFRP specimens
subjected to tensile testing [45, 46]. Similarly, Kempf et al. utilised fourier-transforms, and a
clustering method, based on weighted peak-frequency and partial-powers to differentiate between
matrix-, interphase-, and fibre-failure in the fatigue testing of CFRP composite specimens [47].
Bak and Kalaichalvan also utilised peak-frequency to differentiate between adhesive-, fibre-tear-,
and light fibre-tear-failure in adhesively bonded GFRP lap-joints [48]. An investigation of mode-I
testing of metal-to-metal and metal-to-composite specimens by Droubi et al. also utilised fre-
quency analysis in the form of partial powers, though in this case it was not used to differentiate
between different failure modes, but used to illustrate the variation in frequency as the crack-front
moved through the specimen, changing the source-to-sensor propagation distance [49].
25 of 199
Chapter 2 Literature Review
2.2.7 Time-Frequency-Domain Analysis
Frequency-domain analysis is useful, but does not give the full picture of how the frequency varies
throughout the duration of an AE event. Particularly in cases where dispersive wave types are dom-
inant, the frequency content will vary significantly throughout the event’s duration. A variety of
approaches exist for conversion of time-domain data into time-frequency-domain. The Short-Time
Fourier-Transform (ST-FT) method, also known as Short-Time Fast-Fourier-Transform (ST-FFT),
captures variation in sinusoidal frequency content with time by conducting a Fourier transform of
the signal over multiple short windows of the signal, thus generating a frequency spectra for each
time-window. The use of a fixed time-window does however have the drawback of either poor
frequency-resolution, if the chosen time window is too short, or poor time-resolution if the cho-
sen window is too long. This issue can be avoided by the use of a continuous wavelet-transform
instead of the Fourier transform. The wavelet-transform is based on comparison of the signal
against a short wavelet function, which is scaled and time-shifted, instead of a sinusoidal function
as is used in the Fourier transform. This approach allows the wavelet-transform to effectively use
a longer window at lower frequencies and a shorter window at high frequencies, which ensures
good frequency resolution at low frequency and good time-resolution at high frequency. While a
fully-detailed description of the method is outside the scope of this project, a full description of
the method applied to acoustic emission signals can be found in the work of Suzuki et al. [50].
Wavelet-transforms can theoretically be performed with a variety of different wavelets, including
Gabor (also called Morlet), Meyer, Mexican Hat and Daubechies, but works in acoustic emission
have typically used the Gabor-type wavelet as this provides the best combination of time and fre-
quency resolution of all the available wavelets [51,52]. This is the wavelet-type used by the popu-
lar Vallen Wavelet software [53] used in this work, and also implemented in MATLAB. While the
ST-FT and wavelet-transform have been the most widely used time-frequency transforms for AE
so far, the Choi-Williams transform has also been used by some authors. The Choi-Williams trans-
form provides higher time-frequency resolution than a wavelet-transform, but can be subject to the
generation of parasitic interference in between the original frequency-components [54]. The Choi-
Williams transform has been used most prolifically by Hamstad et al., for a variety of applications,
including investigation of anisotropic attenuation of flexural wave-modes in carbon-fibre compos-
ites [55], the effects of fluid interaction with wave-propagation in multilayered vessels [56], and
a comparison between the Choi-Williams Transform and Wavelet Transform for determination of
group-velocities in aluminium sheets [57].
2.2.8 Modal Analysis
Modal AE analysis may be conducted through a variety of techniques, but is concerned with
analysis of the different wave-modes propagating through the specimen. It can be used when
dispersive wave-types such as Lamb waves exist, leading to separation of the wave-modes in the
time- or time-frequency-domains, and is thus typically suitable for use in sheet-, plate-, bar-, pipe-
or shell-type specimens, as opposed to bulk-type specimens. The general aim of modal analysis is
to relate the wave-modes which are excited, or the relative proportions of the wave-modes excited,
to features of the AE source.
The most basic method of modal analysis is peak-amplitude analysis of the time-domain sig-
nal, with peaks relating to the different wave-modes being identified by comparison with the theo-
retical velocities of the wave-modes. One of the most notable early examples of this is the work of
Gorman, who investigated the effects of source-orientation on plate-wave propagation, by apply-
26 of 199
Chapter 2 Literature Review
ing a PLB source to an aluminium plate at varying angles [58]. It was seen that the early part of
the out-of-plane time-domain signal, corresponding to the extensional mode, reduced in amplitude
with increasing source angle, while the later part of the signal, corresponding to the flexural mode,
increased. To quantify this result, peak-amplitudes of the two modes were taken.
Modal analysis in the time-domain can be effective, but becomes difficult with higher numbers
of wave-modes, overlapping wave-modes or the effects of features such as reflections which can
make it impossible to separate the signals in the time-domain. The solution to this therefore lies
in modal analysis in the time-frequency-domain. Modal analysis in the time-frequency-domain
allows modes which overlap in the time-domain to be separated by frequency. In order to identify
the wave-modes within the time-frequency-domain it is possible to overlay the theoretical Lamb
wave dispersion-curves, providing theoretical arrival times for each wave-mode across the full
frequency-range. This method is well illustrated in work by Hamstad et al. investigating the
use of modal analysis to identify the effects of different source-types and source-depths using a
dynamic finite element simulation [59]. The approach taken by Hamstad et al. to quantify their
modal results was to identify regions of interest relating to each wave-mode, typically regions
of the time-frequency-domain featuring a high level of activity. The peak values of the wavelet-
transform coefficients were then taken from these regions, and the ratio of the coefficients used as
an identifying feature.
Modal AE analysis has a large variety of potential applications. So far it has been applied to
investigation of source orientation [58–60], source depth in simple plate and bar specimens [59,60]
and in the complex geometry of a section of rail track [61]. It has also been used to great effect to
differentiate between different failure modes of fibre-reinforced composites, with it being found
that the in-plane nature of fibre-breakage causes a relatively large symmetric mode, while the
out-of-plane nature of delamination results in a relatively large asymmetric mode [62,63]. Modal
analysis has also been used by Ebrahimkhanlou et al. to develop a single-sensor 2D source-
location method, based on identifying the arrivals of not only the first symmetric and asymmetric
waves, but also their reflections from the edges of the specimens [64,65].
2.2.9 Source-Location
One distinct advantage of AE over other NDT techniques is the ability to accurately locate the
position of damage occurring in real time, by using a network of sensors distributed across the
object under investigation. A variety of methods for achieving source-location are available, and
their suitability depends on the nature of the object under consideration, the required accuracy and
the available equipment.
The simplest source-location method is zonal location. This utilises sensors spaced evenly
across the object at relatively large distances apart (the distance possible will depend on the at-
tenuation of the object). In this method, an AE hit may only be detected by one sensor and can
therefore be assumed to come from within the zone surrounding that sensor. Alternatively, if the
hit is detected at multiple sensors, it can be assigned to the zone surrounding the sensor registering
the highest energy. This method is of low accuracy compared to others, but can be appropriate
for large structures and in situations where accuracy is not highly critical. It has the advantages of
requiring a low number of sensors for the area and also of requiring minimal prior data regarding
the wave-propagation characteristics of the object under test.
Time Difference of Arrival (TDOA) methods provide significantly greater accuracy and can
27 of 199
Chapter 2 Literature Review
assign hits to a point, rather than just a zone. TDOA relies on the use of multiple sensors, two for
one-dimensional (linear) source-location or three for two-dimensional source-location. By using
the known spacing between the sensors and the velocity of wave-propagation in the specimen, the
difference between the arrival times at the different sensors can be used to calculate the source
position. The equations for linear TDOA source location is given below [66]:
l1=1
2(t1t2).v =1
2t.v (2.19)
l2=1
2ll1=1
2(lt.v)(2.20)
Where:
l: The distance between the two sensors.
l1: The distance from the source to the midpoint between the two sensors.
l2: The distance from sensor 1 to the source.
t1,2... : The arrival time of the signal at sensor 1, 2... etc.
v: The wave velocity.
TDOA methods rely on both accurate knowledge of the wave-velocity within the object under
test, and accurate determination of arrival times, both of which can present issues. In the case of
thin sheet or plate specimens, Lamb waves will develop, in which there may be multiple modes
and also multiple frequency-components propagating at different velocities. The use of a simple
threshold-crossing in the time-domain can therefore be insufficient if the frequency and wave-
mode are not known with certainty. More advanced methods, such as that proposed by Hamstad
et al. [67], utilise the wavelet-transform of the signal to identify the arrival times of a selected
mode and frequency, and then use the appropriate corresponding wave-velocity to calculate the
source-location. While this method is suitable for a large number of structure types it assumes a
uniform propagation path and cannot account for features such as holes or changes in thickness,
which can result in reflections of the signal or changes in velocity.
A more appropriate method for geometrically complex specimens is the delta-T method pro-
posed by Baxter et al. [68], in which a simulated source, such as a Hsu-Nielsen source, is applied
at multiple locations across the specimen to generate reference signals. The difference in arrival
times between pairs of sensors for these reference signals can then be mapped, and the resolution
of the map improved by linear interpolation between the tested points. During testing, the sources
can be located by comparison of the recorded differences in arrival times with those of the refer-
ence signals. While the requirement for generation of a database of reference signals makes this
method time-consuming and thus potentially unsuitable for particularly large specimens, it does
have the advantages of requiring no prior assumption of the wave-propagation within the specimen
and is therefore suitable for complex materials and geometries, and also, assuming the sensors are
not moved between generation of the reference signals and testing, no knowledge of the sensor
locations is required.
Another method which can be used to either reduce the number of sensors required, or to
increase the accuracy of TDOA methods, is the use of Single-Sensor-Modal-Analysis-Location
(SSMAL). As previously discussed, the zero-order symmetric and asymmetric components of a
Lamb wave will propagate at different velocities. As demonstrated by Surgeon and Wevers [69],
28 of 199
Chapter 2 Literature Review
and also by Holford and Carter [70], this can be utilised by using the difference in arrival times
between the symmetric and asymmetric modes to calculate the propagation distance from source-
to-sensor. This method does however rely on the propagation of both modes at suitable amplitudes
for the determination of their arrival times, a situation which will not always arise, depending on
the source.
Another method which utilises the principles of SSMAL is the previously mentioned single-
sensor approach based on multimodal edge-reflections taken by Ebrahimkhanlou et al. [64, 65].
This method is proposed for use on isotropic sheet or plate specimens with suitable reflecting
edges. The first stage of the method is to determine the direct source-to-sensor distance, using
the principles of SSMAL, in this case the arrival times being taken from the continuous wavelet-
transform of the signal. The arrival times of the subsequent edge-reflections are then utilised to
determine the distance from the source to the edges of the specimen and to then triangulate the
source-location.
2.3 Artificial Intelligence in AE Analysis
Rather than relying on classification based on a small handful of parameters to differentiate be-
tween source-types, -locations etc, there are now a great variety of computational techniques which
allow analysis of vast quantities of data, and significantly reduce the necessity of human input for
analysis. While there are a great number of techniques, and multiple variations and evolutions of
each, for methods related to AE they can be broken into two main classifications: Untrained and
Trained.
Untrained systems can be presented with input data from multiple different sources, with no
information about which data correlates to each source. The system then aims to separate the data
into groups or ”clusters” of similar data. The system therefore has to decide how many clusters
are needed, how these clusters should be defined, and then which cluster the data belongs to. Once
the data has been clustered, it is up to the user to find correlations between the clusters and the
sources which they may represent. Trained systems, on the other hand, are supplied with input
data which is already classified, for example results from multiple different known test types. The
system is then ”trained” using this data, which is to say the system carries out an optimisation to
find the best way to mathematically differentiate between the known data sets. The system can
then use this optimised method to match any further data to one of the established classifications.
Untrained methods have the advantage of no prior knowledge being necessary; they can be applied
in situations in which the number of potential sources is entirely unknown and then used to help
identify these sources. They do however present the issue that it is not always clear what the
clusters created are actually related to, and it is down to the user to identify relationships between
the clusters and the sources to which they may relate. Trained networks provide much more robust
results, as it is known which phenomena each cluster relates to, but the generation of suitable
training data can be both difficult to achieve and time-consuming.
For application to AE, the inputs for these methods can include typical AE parameters, such
as amplitude, energy, duration, rise-time, peak frequency etc, but can also include the full signal,
the frequency spectra or even the wavelet-transform coefficients of the signal.
The most widely used untrained method in AE is the K-means clustering method. The basic
procedure is to consider the data points for each hit as existing within a multi-dimensional space,
with as many dimensions as there are parameters being used to describe the hit. A number of initial
29 of 199
Chapter 2 Literature Review
mean points are then identified, corresponding to the number of clusters to be formed; this can be
achieved through a number of methods, but most simply the points can be randomly selected. Each
hit is then assigned to the closest mean. The mean is then recalculated based on the values which
have been assigned to it. This process is repeated until a converged solution is achieved in which
each hit is part of a cluster. The number of clusters can either be determined by the user, based on
the number of expected source mechanisms, or determined by mathematical optimisation, which
may be based on factors such as variation within or between the clusters. K-means clustering has
been successfully used in a large number of AE studies in different fields, including some studies
of adhesive-bonding.
It was used by Prathuru [9] in the investigation of the application of in-plane and out-of plane
PLB sources to adhesively-bonded specimens. Four parameters were used, peak amplitude, dura-
tion, and the ratios of energy and peak amplitude found in the low-frequency (<175 kHz) and
high-frequency (>175 kHz) filtered signals. Use of these four parameters was found to provide
robust classification of in-plane and out-of-plane sources.
Pashmforoush et al. [71] utilised a hybrid K-means genetic algorithm for differentiation be-
tween failure modes of core failure, adhesive-bond failure, matrix cracking and fibre-breakage,
during Mode-I testing of composite sandwich panels. The approach taken in this case was to
identify the potential failure modes prior to testing, and to establish the AE characteristics of each
failure type in terms of amplitude and frequency spectra. Following the Mode-I tests, the hy-
brid K-means genetic algorithm was used for clustering. The hybrid method was utilised as it
removes the dependency of the clusters on the initial estimate and prevents the algorithm from
getting stuck in local minima. The clustering was based on the parameters of amplitude, energy
and frequency. The number of appropriate classes was investigated using the Davies-Bouldin in-
dex, which indicated an optimum of four classes, thus corresponding to the number of potential
failure mechanisms identified. The clusters were then assigned to the failure mechanisms based
on the prior investigation of the characteristics of each of the failure modes. The prevalence of
these failure modes in each specimen was then confirmed with SEM imagery. A great number
of other studies have also applied similar methods for clustering of results from fibre-reinforced
composites, typically focusing on input parameters of energy, amplitude, peak-frequency, duration
and rise-time [72–75].
Destousse et al. utilised clustering to aid in analysis of bi-axial loading of scarf joints formed
at a variety of angles [76]. In this case only two parameters were used, peak-frequency and ampli-
tude, and the the number of classes was not pre-specified. A combination of the Davies-Bouldin
index and silhouette coefficient were used to determine the optimum number of clusters, which
was found to be four. At the time of publishing however, no explanation was provided for the
differences between these clusters, although microscopy to further investigate the sources was
claimed to be underway.
A large variety of methods now exist for trained systems to classify data, most of which are
based on the principles of Artificial Neural Networks (ANNs), also known simply as ”Neural
Networks”. A neural network consists of many small units called neurons, which are arranged
into a number of layers, as illustrated in Figure 2.22. The first layer is the input layer, which
contains the numerical inputs, and the final layer is the output layer. Neurons from one layer
are connected to the neurons in the next layer through weighted connections, connections with a
real-valued weight attached to them. In a fully connected network every neuron in one layer is
connected to every neuron in the next layer, while in other more specialised architectures, only
certain combinations of neurons will be connected. The values of the neurons in one layer are
30 of 199
Chapter 2 Literature Review
multiplied by the values of their weighted connections. The bias values of the neurons in the next
layer are then calculated by adding up these values for all of their connections. An activation
function, such as the logistic function y= 1/(1 + e(bias)), is then used to transform the bias
function into a value for the neuron, which can then be passed onto the next layer. This process
is continued until the input values have been propagated through all of the layers to the output.
For a neural network to provide the desired output from a certain set of inputs, the weights of
the connections must be set correctly. The process of setting the weights is know as training the
network, and is achieved using a back-propagation algorithm.
Figure 2.22: Example of a basic fully-connected ANN structure
Artificial neural networks have been used in a wide range of AE applications, including dif-
ferentiation between failure modes in composites [77, 78], source-location [79, 80], detection of
burn- and chatter-faults during grinding processes [81,82], tool wear [83] and detection of partial
discharge in electrical transformers [84].
Kumar et al. [78] tested two different network types, a Radial Basis Function Neural Network
(RBFNN) and a Generalised Regression Neural Network (GRNN), to predict the final failure
strength of sea-water aged GFRP specimens subject to three-point bending. 20 specimens were
used in total, with 16 being used for training and 4 for testing. These were split equally between
ageing periods of 4, 5, 6 and 7 months. The AE parameters used to train the networks were number
of hits, cumulative counts, cumulative energy, cumulative absolute energy and cumulative signal
strength. Analysis of the trained network identified that the network was best able to predict the
final failure strength of a specimen during the period from 500 ms to 800 ms into the test, with
final failure occurring in the range of 1400 ms to 1600 ms. To test the networks, the trained
networks were used to analyse the acoustic emission from the 4 specimens not included in the
training data-set. It was found that within the time region of 500 ms to 800 ms, the networks
could predict the final failure strengths with an error of only 0.5% - 7.2% for the RBFNN type and
0.5% - 4.4% for the GRNN type.
Kalafat and Sause [79] have presented a source localisation method based on the Delta-T
method, but utilising a neural network with the aim of increased accuracy. The experimental setup
used consisted of a cylindrical CFRP pressure vessel with a metallic lining, which was tested
both while empty and whilst filled with water. Seven AE sensors were mounted on the pressure
vessel. A PLB source and a piezoelectric pulser were both used as input sources to cover a wide
range of input frequencies. These were both applied at 444 different points which were marked
out on the vessel. The time difference of arrival between all of the sensor pairs was used as the
31 of 199
Chapter 2 Literature Review
input for the neural network, with the source x and y coordinates as the target output. 111 of the
444 source locations were used to train the network, while the other 333 were used to test it’s
accuracy. Comparison with locations determined using the classical Delta-T method, using the
same test data, showed the neural network based approach to improve source location accuracy by
up to a factor of 6, with the significantly better performance being most prominent in the region
directly in between two sensors, where the time difference of arrival is at its lowest.
Caprino et al [80] have also experimented with the use of neural networks to aid in source lo-
cation, but with a rather different approach. This study focuses on source location in an anisotropic
unidirectional carbon fibre plate. The test setup is fairly typical, with a square 280 mm x 280 mm
CFRP plate, with three sensors located near the edges, and a PLB being used as a source. As in the
work described above by Kalafat and Sause [79], the input to the network is the time difference
of arrival between the sensors, and the target outputs are the coordinates of the source location.
In this case however, the network is not trained with experimental data. A small series of PLB
tests were done to derive an expression for variation in wave velocity with regard to propagation
direction. This was then used to calculate the theoretical time difference of arrival at the sensors
for 2500 randomly selected locations on the plate. These theoretical values were then used as a
training data-set for the neural network. The system was then tested with experimental data from
14 randomly selected locations and was found to perform well, with a mean error of only 2.18
mm. While this approach has been proven to be effective, it is questionable what the advantage
is compared to traditional TDOA methods, as it is still dependent on the accurate determination
of the wave velocity in multiple directions, while systems trained with experimental data have no
reliance on this and are therefore able to cope with complex geometry and variable wave velocities
throughout the specimen.
Kwak and Ha [82] have used a neural network combining inputs from an AE sensor and a
power-meter connected to a grinder to detect and differentiate between burning and chatter vibra-
tion during the grinding process. The network utilises the static and dynamic power parameters
from the power meter, and peak RMS amplitude and Peak Frequency from the AE sensor as inputs.
The simple feed-forward network with two hidden layers was trained using a dataset of 12 sample
signals. When tested using different signal samples, the network identified the faults correctly
with an accuracy of around 95%.
Assessment of machine tool wear by AE has also been achieved by Jemielniak et al [83].
Various AE parameters were tested as inputs for the system, along with feed speed, cutting speed
and cutting forces, which were measured using other sensors. The parameters which were found to
be most useful, and which were used as inputs to the network were the average RMS value of the
signal, and the burst rate, defined as the number of times the RMS value exceeds a preset threshold
in a given time-frame. These parameters were utilised as inputs to a feed-forward network, with a
single output of crater size, a parameter which is directly indicative of tool wear.
A somewhat different approach has been taken by Boczar et al [84] in their investigation of
recognising partial discharges in electrical insulation systems of power transformers. In this case,
instead of using a small number of parameters which summarise the AE signal, such as peak values
or mean values, the approach has been taken of using a much larger input vector. Two different
inputs have been tested, the first being the power spectral density, and the second being the Short
Time Fourier Transform (STFT), with the STFT being rearranged into a single vector with an
arbitrary order, instead of a 2D matrix arranged by frequency and time. A feed forward network
with a single hidden layer of sigmoid neurons was used. Sensitivity studies were carried out to
investigate the effects of the input vector size, with the length of the PSD being varied from 16 to
32 of 199
Chapter 2 Literature Review
1024, and the time interval for the STFT varied from 0.05 ms to 1.95 ms, and also to determine
the most appropriate number of neurons in the hidden layer. Based on a compromise between
accuracy and efficiency (training time), the two final networks produced utilised a PSD length
of 128 or an STFT created with a time interval of 0.4 ms as the input, and used a hidden layer
comprising of 45 neurons. Both of these networks were found to be able to classify the partial
discharges into 8 categories with an accuracy exceeding 95%.
It can be seen from the brief descriptions of the works above, that not only do neural networks
have a lot of potential in AE, in terms of the practical applications for which they may be useful,
but also there is a vast scope for experimentation, in terms of network architecture and the ways
that AE data and features can be utilised as inputs.
2.4 Application of AE to Adhesive-Bonds
The application of AE to adhesively-bonded joints has been explored by a variety of researchers,
with varying levels of insight gained from its use. A selection of these works are described in the
following section and are summarised in Tables 2.1 and 2.2.
Table 2.1: Literature on Application of Acoustic Emission to Adhesive-Bonds
Adhesive Adherend Test Variables Analysis techniques Ref.
Epoxy
FM300-2M
Graphite-
Epoxy
composite
DCB, ENF,
single-lap-
tension
Fracture-mode
Peak-frequency,
pattern-
recognition
[85]
Epoxy
FM300 Composite Single-
lap-shear
Void-size,
Adhesive-thickness,
aging
Acousto-ultrasonic-
parameter [86]
Polyurethane-
SikaForce
7851
GFRP Double-
lap-shear Adhesive thickness Cumulative energy [87]
Epoxy CFRP DCB Temperature
Cumulative energy,
source-location,
clustering
[88]
JGN-T CFRP/Steel Uniaxial-
tension - Cumulative energy [89]
Terokal
5045
CFRP,
GFRP
Single-lap-
shear creep
Temperature,
moisture Cumulative counts [90]
Epoxy-
Araldite
LY556
GFRP
Repair-
patch
-Tension
Repair-patch
type Cumulative counts [91]
Epoxy GFRP DCB,
MMB Fracture-mode Amplitude [92]
33 of 199
Chapter 2 Literature Review
Table 2.2: Literature on Application of Acoustic Emission to Adhesive-Bonds (contd.)
Epoxy GFRP
Single-
lap-shear,
Double-
lap-shear
Adhesive-thickness,
failure-mode
Amplitude,
duration,
peak-frequency,
cumulative counts,
source-location
[48, 93, 94]
Loctite 638 18NiCrMo3
steel
Conical-
torsion Defect-density Cumulative counts [95, 96]
Redux 775,
BSL 308
Aluminium
alloy
Single-
lap-shear
Adhesive-type,
adhesive-thickness Cumulative energy [97]
Loctite AA326
Loctite EA3430
Aluminium
alloy,
CFRP
DCB,
3-ENF
Fracture-mode,
adhesive-type,
defect-density
AE amplitude,
partial powers [8]
Loctite AA326
Loctite EA3430
Aluminium
alloy MMB
Mode-mixity,
adhesive-type,
defect-density
AE energy [98]
Epoxy ER331, Aluminium
alloy
Single-
lap-shear Surface treatment
PCA,
k-means clustering,
cumulative counts
[99]
Epoxy EPG 2601 CFRP
DCB
ENF
ECT
Fracture-mode Amplitude [100]
Epoxy SR150
GFRP-
polyethylene-
foam-
sandwich
DCB Failure-mode
PCA,
K-means clustering,
cumulative counts
[71]
Loctite 326,
Loctite 3430
Aluminium
alloy
PLB,
4-Point-Bend,
Indentation,
Single-
Lap-Shear
Fracture-mode,
Defect density,
Adhesive thickness,
Adhesive type
WT,
modal analysis,
AE energy,
PCA,
k-means clustering
[9]
Araldite
2021,
Sikasil
SG500
Aluminium
alloy DCB Adhesive type
TDOA source-
location,
FEA
[101]
Araldite
LY 1564 GFRP DCB,
3-ENF Fracture-mode
AE Amplitude,
cumulative counts,
cumulative energy
[102]
34 of 199
Chapter 2 Literature Review
The earliest and most basic use of AE for adhesives has been its use as an indicator of the
onset of failure during destructive tests. A large number of studies have used AE-instrumented
destructive tests of adhesively-bonded specimens, and have utilised either amplitude, energy (per
hit), cumulative-energy or cumulative counts to identify the initiation and progression of failure
[8, 71, 87–93, 95–103]. All of these studies have observed that the onset of acoustic emission
correlates well with the load-displacement curves acquired in destructive tests. Typically a low
level of AE activity will slightly precede the drop in load which occurs due to failure, and a
high level of AE activity will be observed during the failure. These types of relationship can be
seen in Figure 2.23. This basic relationship allows AE to be used to successfully detect damage
occurring in an adhesive-bond, and in some cases, to pre-empt total failure of the specimen. This
type of basic study however does not yield the full depth of results that more detailed AE analysis
can provide. More detailed analysis has been utilised not just to identify damage, but to locate
it, to differentiate between different failure modes, different fracture-modes and to estimate the
final failure load. Investigations have also looked at the effects of varying adhesive and adherend
materials, thicknesses and loadings on AE generation.
Figure 2.23: AE amplitude and loading curve for a DCB test performed by Droubi et al. [8]
The earliest use of AE with adhesives is believed to be by Curtis in 1975 [97]. In this study
lap-shear tests were conducted on specimens bonded with a brittle Redux 775 adhesive and a
more ductile BSL 308. The bond strength was varied by varying the thickness of the Redux 775
adhesive, whose strength was noted to be approximately inversely proportional to its thickness.
Cumulative AE energy was investigated as the main parameter of interest, and it was found that
in both adhesives, weaker bonds produced a greater cumulative energy during failure. It was
proposed that, for the Redux 775 adhesive, the cumulative AE energy at fracture (per unit glue-
line volume) was inversely proportional to the lap-shear strength raised to the power n, where n
was less than 6, but this required further analysis to be more clearly defined.
The idea of relating cumulative AE enery to final failure-load was also pursued more recently
by Croccolo and Cuppini [95,96], who applied a similar method of utilising cumulative AE activity
to predict the final releasing moment of an adhesively-bonded conical torsion-test specimen. As
opposed to cumulative energy, cumulative counts of AE were utilised. Testing specimens with
various bond qualities, modified by oiling parts of the adhesive-bond surfaces, revealed that the
gradient of cumulative counts vs applied load corresponded to the treatment of the surfaces, and
therefore to the overall bond strength. It was therefore proposed that a methodology of applying a
low load (25% of the predicted releasing moment), and recording the cumulative AE counts, could
be used to estimate the defective bond area and the total strength of the joint.
35 of 199
Chapter 2 Literature Review
Droubi et al. [98] investigated the Mode-I and -II failure of bonded metal-to-metal and metal-
to-composite specimens through use of AE-instrumented double-cantilever-beam (DCB) and three-
point end-notch-flexure (ENF) tests. Both a ductile and a brittle adhesive were investigated with
varying levels of bond quality, introduced by use of polytetrafluoroethylene (PTFE) spray to re-
duce the effective bond area. As well as noting correspondence between AE activity and features
in the load-curves, it was also recognised - during both calibration tests using a PLB and during
debonding - that there was an increase in both AE amplitude and in the proportion of higher-
frequency spectral content as the source moved closer to the sensor. Analysis was conducted by
Fast Fourier Transform (FFT) and by energy content after band-pass filtering into low-, medium-
and high-frequency-ranges, and therefore considered the entire hit, including the multiple edge-
reflections likely in small specimens.
Liu et al. [102] have also investigated Mode-I and -II failure of adhesive joints using DCB
and ENF tests, but in this case using composite adherends. Very basic AE analysis of cumulative
counts, energy and amplitude was combined with SEM imagery. It was shown that the Mode-I
tests produced a lower number of hits overall, but a higher number of high-amplitude hits than
the Mode-II tests. SEM imagery was used to identify adhesion failure, cohesive failure and fibre-
breakage. It was suggested that the low-amplitude hits corresponded to micro-cracking, while the
high-amplitude hits corresponded to fibre-failure, but no attempt was made to directly relate AE
parameters to adhesive or cohesive failure.
Mode-I fracture was also investigated with a double-cantilever-beam test by Manterola et
al. [101] who focused on the use of TDOA source-location in DCB specimens prepared with both
rigid adhesives and flexible adhesives. Comparison of the AE location with the visually observed
crack-front revealed good correlation in the specimens prepared with rigid adhesives. However,
when the flexible adhesive was used, the AE locations progressed at the same rate as the visual
crack, but AE locations were in the region of 25 mm ahead of the visual crack-tip. Static FEA, us-
ing a bilinear cohesive-zone-model, was used for comparison of the fracture-process zones (FPZ).
It was demonstrated that the flexible adhesive yielded a far greater FPZ than the brittle adhesive.
It was therefore concluded that the AE events originated from the leading edge of the FPZ, rather
than from the visible crack-front as may have ben previously assumed.
One of the features of adhesive-joint failure investigated most successfully using AE is the
failure mode. In metallic specimens, this tends to be differentiation between debonding between
the adherend and adhesive, and cracking of the adhesive itself. In composite specimens the number
of potential failure mechanisms increases significantly, with fibre-breakage, fibre-pullout, fibre-
tear and matrix cracking potentially occurring on top of the two previously mentioned modes.
Bak and Kalaichelvan have conducted a number of studies [48,93,94] in which they have man-
aged to differentiate between failure mechanisms of fibre-tear, light fibre-tear and adhesive failure
by analysis of the peak frequencies of each hit during lap-shear testing of glass-fibre composite
and pure resin single- and double-lap-joint specimens. Comparison of the AE peak frequencies
with scanning electron microscope images allowed identification of correspondence between the
failure-mechanisms and peak-frequencies. The use of a second sensor also allowed for linear
source-location of each hit for further validation of the failure-mechanisms. It was found that
peak-frequencies of <100 kHz corresponded to adhesive failure, 100 kH z to 200 kHz corre-
sponded to light fibre-tear failure and >200 kHz corresponded to fibre-tear failure. While this
method appears successful in the small specimens tested (25.4 mm square bond area), where the
source-sensor distance experiences minimal variation, the results presented by Droubi et al. [98]
indicate that changes in source-sensor separation may lead to changes in spectral content, and
36 of 199
Chapter 2 Literature Review
thus in the failure-mechanism recognised. For application of this method to larger bond areas the
effects of propagation distance may need to be accounted for to ensure reliability of results.
Galy et al. [99] have also successfully differentiated between failure-mechanisms of alu-
minium and epoxy lap-shear specimens. By use of the k-means clustering method, with inputs of
temporal features including amplitude, energy, duration and rise-time, it was possible to attribute
AE hits to either debonding between the adhesive and adherends, or cracking of the adhesive. As
with most of the works relating to AE testing of adhesive-bonds, the bonded area of the specimens
used was of a standard size for a lap-shear test (25 12.5 mm), resulting in minimal variation in
propagation distance. As in Bak and Kalaichelvan [48], the application of this method to larger
specimens - in which propagation distances will vary more significantly - should be approached
cautiously as the dispersion of AE waves with increasing propagation-distance may lead to re-
duced amplitude and energy, and variation in duration and rise-time, while edge-reflections will
also play a more complex role in affecting these factors, dependent on the geometry.
Pashmforoush et al. [71] also utilised a k-means clustering scheme to differentiate between
core-failure, adhesive-debonding, matrix-cracking and fibre-breakage in a Mode-I delamination
test of a composite sandwich structure. In this case, the clustering was done using a hybrid k-
means-genetic algorithm, which provides a more robust clustering method than pure k-means
and is less likely to suffer from becoming stuck in local minima when attempting to establish
the correct number of clusters. The input to the k-means-genetic algorithm was derived from a
principal components analysis of frequency, amplitude and energy. Clusters were defined through
the k-means algorithm, and then attributed to their associated failure mechanisms by comparison
with experimental data from tests in which the failure mechanisms were isolated. The results
of the AE study were then validated with scanning electron microscopy of the damaged areas,
which allowed the failure mechanisms to be observed. Out of the parameters included in the
clustering method, it was observed that frequency provided the best discrimination between the
failure modes, with dominant frequency-ranges of 35 kHz to 65 kH z, 100 kHz to 130 kH z, 170
kHz to 250 kHz and 350 kH z to 450 kHz corresponding to core-failure, adhesive-bond failure,
matrix cracking and fibre-breakage respectively.
Differentiation between fracture-modes (Mode-I = crack-opening and Mode-II = shear) has
also been achieved by acoustic emission through various methods. Dzenis and Saunders con-
ducted AE instrumented Mode-I, -II and mixed mode DCB, ENF and Lap-shear tests. It was
observed that in typical parametric AE analysis, the different tests yielded very similar results,
with hits having a similar amplitude and similar frequency spectra. However, using the Vallen
software VisualClass it was possible to differentiate the signals using statistical pattern recogni-
tion. VisualClass normalises the waveform, and then breaks it into smaller time-windows. The
FFT of each of those time-windows is then taken, resulting in a spectrum of a number of data-
points. These values, along with the normalisation factors, are then used as features. Visual-class
then selects the features of highest discrimination quality, based on a set of training data. In this
case the results from the DCB and ENF tests were used as the training data. It was found that
by use of this approach the pure Mode-I and -II signals could be fully separated in feature-feature
space. Using the same discriminating features to analyse the results of the lap-shear test resulted
in a data cluster which overlapped significantly with the Mode-II results, which is as should be
expected considering that a lap-shear test is predominantly a Mode-II test. This approach of utilis-
ing statistical pattern recognition has proven to be extremely powerful and able to succeed where
parametric analysis fails, and has demonstrated that there is a significant difference in the signals
acquired from Mode-I and Mode-II fracture. It does not however shed any light onto what the
37 of 199
Chapter 2 Literature Review
actual differences between these signals are.
Prathuru [9] takes a rather different approach to differentiating between Mode-I and Mode-II
type hits during four-point flexure and lap-shear testing. PLB tests are carried out on both the face
(in-plane) and the end (out-of-plane) of various bonded specimens. It is proposed that Mode-I
fracture will correspond to an out-of-plane source, while Mode-II fracture will correspond better
to an in-plane source. K-means clustering is then used to differentiate between the signals and to
identify the parameters that give greatest discrimination between the two source orientations. The
parameter identified as having the greatest potential for discrimination being the ratio between
energy in the high ( >175 kHz) and low ( <175 kHz) frequency bands. This clustering method
was then applied to the data from a four-point-bending test and a lap-shear test. It was found that
both tests produced both Mode-I and -II type signals, but were dominated by the Mode-II type. As
four-point-bending and lap-shear tests are both predominantly Mode-II tests, this is the expected
result. While this attempt to differentiate between fracture-modes appears to be successful, it
is difficult to assess the true validity of this method as Mode-I dominant fracture tests were not
carried out. It could therefore be the case that adhesive failure of any type corresponds better to
the in-plane PLB than the out-of-plane PLB, due to the differences in source material (graphite vs
adhesive material) and source-type and location (surface monopole vs buried dipole).
2.5 Finite Element Simulation of Acoustic Emission
Numerical modelling and simulation of acoustic emission can provide a greater understanding of
the underlying physics, and provides the ability to quickly and cost-effectively investigate the ef-
fects of multiple different parameters affecting the generation, propagation and detection of AE,
in an environment free from the sources of error and variation typically present in an experimen-
tal set-up. Numerical modelling of AE does however have the disadvantage of being incredibly
computationally expensive, due to the requirements for a well refined mesh, and for a high num-
ber of small time-steps, thus requiring significant RAM, memory for large volumes of temporary
files, and a suitably fast processor. With the continued rapid development of the computer indus-
try, these disadvantages are however likely to become significantly less of a challenge in future.
While there is no substitute for real-world experimental data for studies of different AE sources,
FE simulations are superior to experimental results in terms of exact knowledge of their source-
location, size, magnitude and orientation. The option to record the absolute surface displacement
or velocity without the significant effects of the sensor type used (although this can be introduced
in a totally controlled manner if desired); the lack of noise and the ability to include or exclude
reflections as is deemed appropriate [59]. AE propagation is relatively well understood and draws
on knowledge established in the field of ultrasonics, as well as work conducted in seismology. The
processes involved in the generation of AE sources is however far less well documented, as is the
detection of AE signals. A variety of methods of varying complexity have been implemented by
previous authors to simulate AE generation and detection. This section will give an overview of
some of the key literature in this field, as summarised in Table 2.3, giving particular emphasis to
the simulation methodology and set-up, as opposed to the results gained.
38 of 199
Chapter 2 Literature Review
Table 2.3: Literature on Finite Element Simulation of Acoustic Emission
Parameter Investigated Source Sensor 2D/3D Software Ref.
Rise-time, source-width,
stress-distribution
and sensor diameter
Surface force,
linear rise and
time-dependent
Surface area 2D
Axi.
NIST
Boulder [104]
Source size, rise-time
and mesh resolution
effects on 1” steel plate
Surface force,
time-dependent
Point, high-
pass filtered 3D NIST
Boulder [105]
PLB source modelled
on isotropic and
anisotropic sheet specimens
Surface force,
linear ramp Point 2D Axi.
3D
NIST
Boulder [106]
Edge-reflections in
aluminium plate specimens,
in- and out-of-plane sources
Surface force Point, high-
pass filtered 3D NIST
Boulder [107]
Modal analysis of source-type,
-depth, propagation distance
and source-location
on aluminium plates
Cosine-bell
buried dipole,
surface monopole,
Point, high-
pass filtered 3D NIST
Boulder
[59],
[67]
Modal analysis of source rise
time in thin plates
Cosine-bell
buried dipole
Point, high-
pass filtered 2D Axi. NIST
Boulder [108]
Matrix-cracking, fibre-breakage
and fibre-matrix interface
-failure in CFRP.
Buried cross-shape
multi-material
crack source
WD-type
surface sensor 3D COMSOL [109]
Inhomogeneous source
materials to represent
resin or fibre-failure
Buried dipole
Time dependent Point 3D COMSOL [110]
Dipole source
characteristics
Buried dipole
time dependent Point 2d Axi. COMSOL [111]
Detailed modelling of
PLB source fracture Pencil lead fracture Point 2D COMSOL [112]
Effects of holes,
rivets and delaminations
in CFRP sheet
Buried dipole,
linear ramp
Point, high-
pass filtered 3D COMSOL [113]
Investigation of anisotropic
attenuation behaviour of CFRP Cosine-bell force Point 3D ABAQUS [55]
Lamb-mode interaction with
macroscopic CFRP defects Buried dipole - 3D COMSOL [114]
Lamb-mode propagation in
isotropic/anisotropic layered
hybrid composite
Buried dipole - 3D COMSOL [115]
Fibre-failure and
matrix cracking
CZM crack-
propagation
Multi-physics
sensor and
pre-amp
3D COMSOL [116]
Modal analysis of buried
dipole source angles
in PVC rod
Cosine-bell
buried dipole Point 3D NIST
Boulder [117]
Modal analysis investigating
source depth and direction
Buried dipoles,
linear rise-time
Point,
surface with
transfer function
3D ABAQUS [60]
Source-location of PLB
on anisotropic honeycomb
sandwich structure
Time-varying
force Point 3D ABAQUS [118]
Lamb wave generation by
PLB on aluminium sheet
Cosine-bell
surface force Point 2D ANSYS [119]
39 of 199
Chapter 2 Literature Review
Early works by Hamstad, Gary, Prosser and OGallagher [104–106] investigated the FE mod-
elling of out-of-plane and in-plane PLBs on aluminium and steel plate specimens, initially utilis-
ing a simplified axisymmetric 2D finite element code developed by Blake and Bond [120], before
moving to a 3D model. These studies aimed to validate the FE models by comparison with ex-
perimental data for the PLB tests and also by comparison with Mindlin plate theory. Particular
attention was given to the effects of mesh size, source size, source rise-time and source-type. In
these cases the source models used were, for the majority of tests, a uniformly-distributed time-
varying force with a maximum value of 1 N applied over the contact area of the PLB source. A
force of constant intensity and increasing application-area was also trialled and found to create
a similar result. As the actual source from a PLB results from the removal of load as the lead
breaks, as opposed to the deformation as the lead is applied, the results obtained were 180° out
of phase with experimental results, but otherwise accurate. Mesh sizes were varied from 0.6 mm
to 0.023 mm, and time steps chosen to be between 0.082 µs and 0.0033 µs as to satisfy the CFL
stability condition, which requires the time steps to be shorter than the time taken for the high-
est velocity wave to cross an element. These studies successfully demonstrated good correlation
between between experimental, theoretical and FE results and demonstrated the potential of FE
simulations of AE. It was observed that to generate AE in the typical detectable range of over
100 kHz, source times must be of the order of 10 µs or less. Attempts to determine the effect
of source size were somewhat limited by the mesh refinement which could be achieved with the
minimal random access memory of the early computers which were being used. It was however
observed that a source diameter of 0.528 mm or 1.1 mm provided good time-resolution in the
generated Rayleigh waves, while a 4.23 mm diameter source gave poor results, leading the au-
thors to conclude that a source diameter of less than 3 mm should provide a sufficiently accurate
result. The sensor was modelled in these studies as a disc on the surface of the specimen, with the
electrical output of the sensor being assumed to be directly proportional to the average surface dis-
placement over the area of this disc. Comparisons were made between sensor diameters of 2 mm,
3.18 mm, 6.35 mm, 12.7 mm and 25.4 mm, with the two smallest sensors reportedly producing
near identical results to the displacement profile of a single point at the sensor’s centre, while the
6.35 mm sensor still produced relatively accurate, but observably different, results, and the two
larger sensors suffered from significant distortion as the sensor diameter became comparable to
the wavelength. The outputs from these simulated sensors were filtered to a suitable frequency
band for AE using a simple 50 kHz [105] or 100 kHz [104] high-pass filter. No further attempts
were made to model the effects of sensor coupling, frequency response or the effects of the data
acquisition system on the signal. These initial simulations were only run for the first 150 µs of the
signal, and thus included the initial S0and A0or Rayleigh waves, but did not include the effects
of edge-reflections.
Further work by the same group in 1999 [107] introduced 3D simulations of in-plane and
out-of-plane PLB tests conducted on aluminium plates, with source- and sensor-configurations
designed to induce both normal- and oblique-incidence edge-reflections. The same 3D modelling
set-up was used as in previous works, with a time-varying normal force with a maximum value of
1 N applied a 0.3 mm diameter area on the specimen surface. The FE results were compared with
experimental results acquired using an absolutely calibrated sensor with a flat frequency response
from 20 kHz to 1 MHz. Post-processing of signals to allow direct comparison consisted of invert-
ing the signal (to account for the source being 180° out of phase) and filtering both experimental
and FE signals with 50 kHz to 1 MHz or 100 kHz to 1 MHz band-pass filter. The upper and lower
limits were selected due to the limited response of the sensor above 1MHz and below 20 kHz. The
100 kHz limit was then used to remove the low velocity A0mode which obscured S0reflections
40 of 199
Chapter 2 Literature Review
during investigation of in-plane sources. This study found excellent correlation between the ex-
perimental and FE results and demonstrated that 3D FE simulations were capable of accurately
predicting edge-reflections.
The modelling of buried monopole and dipole sources, which are more representative of the
majority of AE sources than a force applied to an external surface, was also introduced by Hamstad
et al. in 1999 [121]. Using the same axisymmetric model as in their previous studies, this study
investigated the effects of source sizes and mesh sizing as well as demonstrating the ability to
model buried monopole and dipole sources in steel plates. The monopole source was created using
a single vertical body force located at the midplane of the plate. The dipole was created with two
simultaneously-opposing body-forces separated by a small distance. In both cases the force was
applied over a small but finite area, thus meaning that the forces were not technically point sources,
but were small enough to be considered as such. The forces were applied as cosine-bell-type step
functions with a rise-time of 0.5 µs. Due to the difficulty in experimentally validating the results
of the simulation they were compared with the analytical solutions of Scruby et al. [122], Pao et
al. [123] and Hsu [124]. The FE solutions showed good agreement with the analytical solutions,
thus validating this method of modelling buried AE sources. The investigation of source size and
mesh size concluded that the ratio between the minimum wavelength of interest and the source size
must be at least two, while the ratio between wavelength and mesh size must be at least fifteen for
adequate results to be obtained.
This work was followed up by a two-part article by Hamstad, OGallagher and Gary, which
investigated the application of wavelet-transforms to a database of AE signals generated by FEA.
The first article [59] focused on source identification, while the second [67] investigated the ap-
plication to source-location. The first investigation utilised 4.7 mm thick aluminium specimens
with lateral dimensions of 1000 mm by 1000 mm and 480 mm by 25.4 mm, giving the potential
to investigate a case analogous to an infinite specimen (edge-reflections can be easily excluded)
and a small coupon-type specimen. The study considered three types of buried point source; a 1
N magnitude in-plane dipole aligned with the direction of propagation towards the sensors, a 1 N
magnitude out-of-plane dipole, and a source designed to represent crack initiation. The crack ini-
tiation source consisted of three dipoles, the largest being a 1 N dipole aligned in the propagation
direction towards the sensors, and dipoles in the other two directions having magnitudes of 0.52
N. The dipoles were each composed of a central cell with the cells either side being subject to
opposing body forces. As in previous works, the sources forces varied with a cosine-bell-type step
function, with a rise-time of 1.5 µs. The mesh consisted of uniform three-dimensional elements of
0.313 mm and time steps of 0.045 µs were used during the simulation. Out-of-plane displacement
of points located 60 mm, 120 mm and 180 mm from the source were taken as the recorded AE
signal, thus obtaining a perfect signal, neglecting the aperture effects of sensor size or any effects
of sensor coupling. To make the simulated results more directly comparable with experimentally
acquired AE signals the results were re-sampled to a step size of 0.1 µs (sampling frequency 10
MHz) and